brintos

brintos / linux-shallow public Read only

0
0
Text · 49.0 KiB · 37a6794 Raw
1906 lines · c
1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause2/*3 * Copyright(c) 2015-2020 Intel Corporation.4 * Copyright(c) 2021 Cornelis Networks.5 */6 7#include <linux/spinlock.h>8#include <linux/pci.h>9#include <linux/io.h>10#include <linux/delay.h>11#include <linux/netdevice.h>12#include <linux/vmalloc.h>13#include <linux/module.h>14#include <linux/prefetch.h>15#include <rdma/ib_verbs.h>16#include <linux/etherdevice.h>17 18#include "hfi.h"19#include "trace.h"20#include "qp.h"21#include "sdma.h"22#include "debugfs.h"23#include "vnic.h"24#include "fault.h"25 26#include "ipoib.h"27#include "netdev.h"28 29#undef pr_fmt30#define pr_fmt(fmt) DRIVER_NAME ": " fmt31 32DEFINE_MUTEX(hfi1_mutex);	/* general driver use */33 34unsigned int hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;35module_param_named(max_mtu, hfi1_max_mtu, uint, S_IRUGO);36MODULE_PARM_DESC(max_mtu, "Set max MTU bytes, default is " __stringify(37		 HFI1_DEFAULT_MAX_MTU));38 39unsigned int hfi1_cu = 1;40module_param_named(cu, hfi1_cu, uint, S_IRUGO);41MODULE_PARM_DESC(cu, "Credit return units");42 43unsigned long hfi1_cap_mask = HFI1_CAP_MASK_DEFAULT;44static int hfi1_caps_set(const char *val, const struct kernel_param *kp);45static int hfi1_caps_get(char *buffer, const struct kernel_param *kp);46static const struct kernel_param_ops cap_ops = {47	.set = hfi1_caps_set,48	.get = hfi1_caps_get49};50module_param_cb(cap_mask, &cap_ops, &hfi1_cap_mask, S_IWUSR | S_IRUGO);51MODULE_PARM_DESC(cap_mask, "Bit mask of enabled/disabled HW features");52 53MODULE_LICENSE("Dual BSD/GPL");54MODULE_DESCRIPTION("Cornelis Omni-Path Express driver");55 56/*57 * MAX_PKT_RCV is the max # if packets processed per receive interrupt.58 */59#define MAX_PKT_RECV 6460/*61 * MAX_PKT_THREAD_RCV is the max # of packets processed before62 * the qp_wait_list queue is flushed.63 */64#define MAX_PKT_RECV_THREAD (MAX_PKT_RECV * 4)65#define EGR_HEAD_UPDATE_THRESHOLD 1666 67struct hfi1_ib_stats hfi1_stats;68 69static int hfi1_caps_set(const char *val, const struct kernel_param *kp)70{71	int ret = 0;72	unsigned long *cap_mask_ptr = (unsigned long *)kp->arg,73		cap_mask = *cap_mask_ptr, value, diff,74		write_mask = ((HFI1_CAP_WRITABLE_MASK << HFI1_CAP_USER_SHIFT) |75			      HFI1_CAP_WRITABLE_MASK);76 77	ret = kstrtoul(val, 0, &value);78	if (ret) {79		pr_warn("Invalid module parameter value for 'cap_mask'\n");80		goto done;81	}82	/* Get the changed bits (except the locked bit) */83	diff = value ^ (cap_mask & ~HFI1_CAP_LOCKED_SMASK);84 85	/* Remove any bits that are not allowed to change after driver load */86	if (HFI1_CAP_LOCKED() && (diff & ~write_mask)) {87		pr_warn("Ignoring non-writable capability bits %#lx\n",88			diff & ~write_mask);89		diff &= write_mask;90	}91 92	/* Mask off any reserved bits */93	diff &= ~HFI1_CAP_RESERVED_MASK;94	/* Clear any previously set and changing bits */95	cap_mask &= ~diff;96	/* Update the bits with the new capability */97	cap_mask |= (value & diff);98	/* Check for any kernel/user restrictions */99	diff = (cap_mask & (HFI1_CAP_MUST_HAVE_KERN << HFI1_CAP_USER_SHIFT)) ^100		((cap_mask & HFI1_CAP_MUST_HAVE_KERN) << HFI1_CAP_USER_SHIFT);101	cap_mask &= ~diff;102	/* Set the bitmask to the final set */103	*cap_mask_ptr = cap_mask;104done:105	return ret;106}107 108static int hfi1_caps_get(char *buffer, const struct kernel_param *kp)109{110	unsigned long cap_mask = *(unsigned long *)kp->arg;111 112	cap_mask &= ~HFI1_CAP_LOCKED_SMASK;113	cap_mask |= ((cap_mask & HFI1_CAP_K2U) << HFI1_CAP_USER_SHIFT);114 115	return sysfs_emit(buffer, "0x%lx\n", cap_mask);116}117 118struct pci_dev *get_pci_dev(struct rvt_dev_info *rdi)119{120	struct hfi1_ibdev *ibdev = container_of(rdi, struct hfi1_ibdev, rdi);121	struct hfi1_devdata *dd = container_of(ibdev,122					       struct hfi1_devdata, verbs_dev);123	return dd->pcidev;124}125 126/*127 * Return count of units with at least one port ACTIVE.128 */129int hfi1_count_active_units(void)130{131	struct hfi1_devdata *dd;132	struct hfi1_pportdata *ppd;133	unsigned long index, flags;134	int pidx, nunits_active = 0;135 136	xa_lock_irqsave(&hfi1_dev_table, flags);137	xa_for_each(&hfi1_dev_table, index, dd) {138		if (!(dd->flags & HFI1_PRESENT) || !dd->kregbase1)139			continue;140		for (pidx = 0; pidx < dd->num_pports; ++pidx) {141			ppd = dd->pport + pidx;142			if (ppd->lid && ppd->linkup) {143				nunits_active++;144				break;145			}146		}147	}148	xa_unlock_irqrestore(&hfi1_dev_table, flags);149	return nunits_active;150}151 152/*153 * Get address of eager buffer from it's index (allocated in chunks, not154 * contiguous).155 */156static inline void *get_egrbuf(const struct hfi1_ctxtdata *rcd, u64 rhf,157			       u8 *update)158{159	u32 idx = rhf_egr_index(rhf), offset = rhf_egr_buf_offset(rhf);160 161	*update |= !(idx & (rcd->egrbufs.threshold - 1)) && !offset;162	return (void *)(((u64)(rcd->egrbufs.rcvtids[idx].addr)) +163			(offset * RCV_BUF_BLOCK_SIZE));164}165 166static inline void *hfi1_get_header(struct hfi1_ctxtdata *rcd,167				    __le32 *rhf_addr)168{169	u32 offset = rhf_hdrq_offset(rhf_to_cpu(rhf_addr));170 171	return (void *)(rhf_addr - rcd->rhf_offset + offset);172}173 174static inline struct ib_header *hfi1_get_msgheader(struct hfi1_ctxtdata *rcd,175						   __le32 *rhf_addr)176{177	return (struct ib_header *)hfi1_get_header(rcd, rhf_addr);178}179 180static inline struct hfi1_16b_header181		*hfi1_get_16B_header(struct hfi1_ctxtdata *rcd,182				     __le32 *rhf_addr)183{184	return (struct hfi1_16b_header *)hfi1_get_header(rcd, rhf_addr);185}186 187/*188 * Validate and encode the a given RcvArray Buffer size.189 * The function will check whether the given size falls within190 * allowed size ranges for the respective type and, optionally,191 * return the proper encoding.192 */193int hfi1_rcvbuf_validate(u32 size, u8 type, u16 *encoded)194{195	if (unlikely(!PAGE_ALIGNED(size)))196		return 0;197	if (unlikely(size < MIN_EAGER_BUFFER))198		return 0;199	if (size >200	    (type == PT_EAGER ? MAX_EAGER_BUFFER : MAX_EXPECTED_BUFFER))201		return 0;202	if (encoded)203		*encoded = ilog2(size / PAGE_SIZE) + 1;204	return 1;205}206 207static void rcv_hdrerr(struct hfi1_ctxtdata *rcd, struct hfi1_pportdata *ppd,208		       struct hfi1_packet *packet)209{210	struct ib_header *rhdr = packet->hdr;211	u32 rte = rhf_rcv_type_err(packet->rhf);212	u32 mlid_base;213	struct hfi1_ibport *ibp = rcd_to_iport(rcd);214	struct hfi1_devdata *dd = ppd->dd;215	struct hfi1_ibdev *verbs_dev = &dd->verbs_dev;216	struct rvt_dev_info *rdi = &verbs_dev->rdi;217 218	if ((packet->rhf & RHF_DC_ERR) &&219	    hfi1_dbg_fault_suppress_err(verbs_dev))220		return;221 222	if (packet->rhf & RHF_ICRC_ERR)223		return;224 225	if (packet->etype == RHF_RCV_TYPE_BYPASS) {226		goto drop;227	} else {228		u8 lnh = ib_get_lnh(rhdr);229 230		mlid_base = be16_to_cpu(IB_MULTICAST_LID_BASE);231		if (lnh == HFI1_LRH_BTH) {232			packet->ohdr = &rhdr->u.oth;233		} else if (lnh == HFI1_LRH_GRH) {234			packet->ohdr = &rhdr->u.l.oth;235			packet->grh = &rhdr->u.l.grh;236		} else {237			goto drop;238		}239	}240 241	if (packet->rhf & RHF_TID_ERR) {242		/* For TIDERR and RC QPs preemptively schedule a NAK */243		u32 tlen = rhf_pkt_len(packet->rhf); /* in bytes */244		u32 dlid = ib_get_dlid(rhdr);245		u32 qp_num;246 247		/* Sanity check packet */248		if (tlen < 24)249			goto drop;250 251		/* Check for GRH */252		if (packet->grh) {253			u32 vtf;254			struct ib_grh *grh = packet->grh;255 256			if (grh->next_hdr != IB_GRH_NEXT_HDR)257				goto drop;258			vtf = be32_to_cpu(grh->version_tclass_flow);259			if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)260				goto drop;261		}262 263		/* Get the destination QP number. */264		qp_num = ib_bth_get_qpn(packet->ohdr);265		if (dlid < mlid_base) {266			struct rvt_qp *qp;267			unsigned long flags;268 269			rcu_read_lock();270			qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);271			if (!qp) {272				rcu_read_unlock();273				goto drop;274			}275 276			/*277			 * Handle only RC QPs - for other QP types drop error278			 * packet.279			 */280			spin_lock_irqsave(&qp->r_lock, flags);281 282			/* Check for valid receive state. */283			if (!(ib_rvt_state_ops[qp->state] &284			      RVT_PROCESS_RECV_OK)) {285				ibp->rvp.n_pkt_drops++;286			}287 288			switch (qp->ibqp.qp_type) {289			case IB_QPT_RC:290				hfi1_rc_hdrerr(rcd, packet, qp);291				break;292			default:293				/* For now don't handle any other QP types */294				break;295			}296 297			spin_unlock_irqrestore(&qp->r_lock, flags);298			rcu_read_unlock();299		} /* Unicast QP */300	} /* Valid packet with TIDErr */301 302	/* handle "RcvTypeErr" flags */303	switch (rte) {304	case RHF_RTE_ERROR_OP_CODE_ERR:305	{306		void *ebuf = NULL;307		u8 opcode;308 309		if (rhf_use_egr_bfr(packet->rhf))310			ebuf = packet->ebuf;311 312		if (!ebuf)313			goto drop; /* this should never happen */314 315		opcode = ib_bth_get_opcode(packet->ohdr);316		if (opcode == IB_OPCODE_CNP) {317			/*318			 * Only in pre-B0 h/w is the CNP_OPCODE handled319			 * via this code path.320			 */321			struct rvt_qp *qp = NULL;322			u32 lqpn, rqpn;323			u16 rlid;324			u8 svc_type, sl, sc5;325 326			sc5 = hfi1_9B_get_sc5(rhdr, packet->rhf);327			sl = ibp->sc_to_sl[sc5];328 329			lqpn = ib_bth_get_qpn(packet->ohdr);330			rcu_read_lock();331			qp = rvt_lookup_qpn(rdi, &ibp->rvp, lqpn);332			if (!qp) {333				rcu_read_unlock();334				goto drop;335			}336 337			switch (qp->ibqp.qp_type) {338			case IB_QPT_UD:339				rlid = 0;340				rqpn = 0;341				svc_type = IB_CC_SVCTYPE_UD;342				break;343			case IB_QPT_UC:344				rlid = ib_get_slid(rhdr);345				rqpn = qp->remote_qpn;346				svc_type = IB_CC_SVCTYPE_UC;347				break;348			default:349				rcu_read_unlock();350				goto drop;351			}352 353			process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);354			rcu_read_unlock();355		}356 357		packet->rhf &= ~RHF_RCV_TYPE_ERR_SMASK;358		break;359	}360	default:361		break;362	}363 364drop:365	return;366}367 368static inline void init_packet(struct hfi1_ctxtdata *rcd,369			       struct hfi1_packet *packet)370{371	packet->rsize = get_hdrqentsize(rcd); /* words */372	packet->maxcnt = get_hdrq_cnt(rcd) * packet->rsize; /* words */373	packet->rcd = rcd;374	packet->updegr = 0;375	packet->etail = -1;376	packet->rhf_addr = get_rhf_addr(rcd);377	packet->rhf = rhf_to_cpu(packet->rhf_addr);378	packet->rhqoff = hfi1_rcd_head(rcd);379	packet->numpkt = 0;380}381 382/* We support only two types - 9B and 16B for now */383static const hfi1_handle_cnp hfi1_handle_cnp_tbl[2] = {384	[HFI1_PKT_TYPE_9B] = &return_cnp,385	[HFI1_PKT_TYPE_16B] = &return_cnp_16B386};387 388/**389 * hfi1_process_ecn_slowpath - Process FECN or BECN bits390 * @qp: The packet's destination QP391 * @pkt: The packet itself.392 * @prescan: Is the caller the RXQ prescan393 *394 * Process the packet's FECN or BECN bits. By now, the packet395 * has already been evaluated whether processing of those bit should396 * be done.397 * The significance of the @prescan argument is that if the caller398 * is the RXQ prescan, a CNP will be send out instead of waiting for the399 * normal packet processing to send an ACK with BECN set (or a CNP).400 */401bool hfi1_process_ecn_slowpath(struct rvt_qp *qp, struct hfi1_packet *pkt,402			       bool prescan)403{404	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);405	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);406	struct ib_other_headers *ohdr = pkt->ohdr;407	struct ib_grh *grh = pkt->grh;408	u32 rqpn = 0;409	u16 pkey;410	u32 rlid, slid, dlid = 0;411	u8 hdr_type, sc, svc_type, opcode;412	bool is_mcast = false, ignore_fecn = false, do_cnp = false,413		fecn, becn;414 415	/* can be called from prescan */416	if (pkt->etype == RHF_RCV_TYPE_BYPASS) {417		pkey = hfi1_16B_get_pkey(pkt->hdr);418		sc = hfi1_16B_get_sc(pkt->hdr);419		dlid = hfi1_16B_get_dlid(pkt->hdr);420		slid = hfi1_16B_get_slid(pkt->hdr);421		is_mcast = hfi1_is_16B_mcast(dlid);422		opcode = ib_bth_get_opcode(ohdr);423		hdr_type = HFI1_PKT_TYPE_16B;424		fecn = hfi1_16B_get_fecn(pkt->hdr);425		becn = hfi1_16B_get_becn(pkt->hdr);426	} else {427		pkey = ib_bth_get_pkey(ohdr);428		sc = hfi1_9B_get_sc5(pkt->hdr, pkt->rhf);429		dlid = qp->ibqp.qp_type != IB_QPT_UD ? ib_get_dlid(pkt->hdr) :430			ppd->lid;431		slid = ib_get_slid(pkt->hdr);432		is_mcast = (dlid > be16_to_cpu(IB_MULTICAST_LID_BASE)) &&433			   (dlid != be16_to_cpu(IB_LID_PERMISSIVE));434		opcode = ib_bth_get_opcode(ohdr);435		hdr_type = HFI1_PKT_TYPE_9B;436		fecn = ib_bth_get_fecn(ohdr);437		becn = ib_bth_get_becn(ohdr);438	}439 440	switch (qp->ibqp.qp_type) {441	case IB_QPT_UD:442		rlid = slid;443		rqpn = ib_get_sqpn(pkt->ohdr);444		svc_type = IB_CC_SVCTYPE_UD;445		break;446	case IB_QPT_SMI:447	case IB_QPT_GSI:448		rlid = slid;449		rqpn = ib_get_sqpn(pkt->ohdr);450		svc_type = IB_CC_SVCTYPE_UD;451		break;452	case IB_QPT_UC:453		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);454		rqpn = qp->remote_qpn;455		svc_type = IB_CC_SVCTYPE_UC;456		break;457	case IB_QPT_RC:458		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);459		rqpn = qp->remote_qpn;460		svc_type = IB_CC_SVCTYPE_RC;461		break;462	default:463		return false;464	}465 466	ignore_fecn = is_mcast || (opcode == IB_OPCODE_CNP) ||467		(opcode == IB_OPCODE_RC_ACKNOWLEDGE);468	/*469	 * ACKNOWLEDGE packets do not get a CNP but this will be470	 * guarded by ignore_fecn above.471	 */472	do_cnp = prescan ||473		(opcode >= IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST &&474		 opcode <= IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE) ||475		opcode == TID_OP(READ_RESP) ||476		opcode == TID_OP(ACK);477 478	/* Call appropriate CNP handler */479	if (!ignore_fecn && do_cnp && fecn)480		hfi1_handle_cnp_tbl[hdr_type](ibp, qp, rqpn, pkey,481					      dlid, rlid, sc, grh);482 483	if (becn) {484		u32 lqpn = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;485		u8 sl = ibp->sc_to_sl[sc];486 487		process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);488	}489	return !ignore_fecn && fecn;490}491 492struct ps_mdata {493	struct hfi1_ctxtdata *rcd;494	u32 rsize;495	u32 maxcnt;496	u32 ps_head;497	u32 ps_tail;498	u32 ps_seq;499};500 501static inline void init_ps_mdata(struct ps_mdata *mdata,502				 struct hfi1_packet *packet)503{504	struct hfi1_ctxtdata *rcd = packet->rcd;505 506	mdata->rcd = rcd;507	mdata->rsize = packet->rsize;508	mdata->maxcnt = packet->maxcnt;509	mdata->ps_head = packet->rhqoff;510 511	if (get_dma_rtail_setting(rcd)) {512		mdata->ps_tail = get_rcvhdrtail(rcd);513		if (rcd->ctxt == HFI1_CTRL_CTXT)514			mdata->ps_seq = hfi1_seq_cnt(rcd);515		else516			mdata->ps_seq = 0; /* not used with DMA_RTAIL */517	} else {518		mdata->ps_tail = 0; /* used only with DMA_RTAIL*/519		mdata->ps_seq = hfi1_seq_cnt(rcd);520	}521}522 523static inline int ps_done(struct ps_mdata *mdata, u64 rhf,524			  struct hfi1_ctxtdata *rcd)525{526	if (get_dma_rtail_setting(rcd))527		return mdata->ps_head == mdata->ps_tail;528	return mdata->ps_seq != rhf_rcv_seq(rhf);529}530 531static inline int ps_skip(struct ps_mdata *mdata, u64 rhf,532			  struct hfi1_ctxtdata *rcd)533{534	/*535	 * Control context can potentially receive an invalid rhf.536	 * Drop such packets.537	 */538	if ((rcd->ctxt == HFI1_CTRL_CTXT) && (mdata->ps_head != mdata->ps_tail))539		return mdata->ps_seq != rhf_rcv_seq(rhf);540 541	return 0;542}543 544static inline void update_ps_mdata(struct ps_mdata *mdata,545				   struct hfi1_ctxtdata *rcd)546{547	mdata->ps_head += mdata->rsize;548	if (mdata->ps_head >= mdata->maxcnt)549		mdata->ps_head = 0;550 551	/* Control context must do seq counting */552	if (!get_dma_rtail_setting(rcd) ||553	    rcd->ctxt == HFI1_CTRL_CTXT)554		mdata->ps_seq = hfi1_seq_incr_wrap(mdata->ps_seq);555}556 557/*558 * prescan_rxq - search through the receive queue looking for packets559 * containing Excplicit Congestion Notifications (FECNs, or BECNs).560 * When an ECN is found, process the Congestion Notification, and toggle561 * it off.562 * This is declared as a macro to allow quick checking of the port to avoid563 * the overhead of a function call if not enabled.564 */565#define prescan_rxq(rcd, packet) \566	do { \567		if (rcd->ppd->cc_prescan) \568			__prescan_rxq(packet); \569	} while (0)570static void __prescan_rxq(struct hfi1_packet *packet)571{572	struct hfi1_ctxtdata *rcd = packet->rcd;573	struct ps_mdata mdata;574 575	init_ps_mdata(&mdata, packet);576 577	while (1) {578		struct hfi1_ibport *ibp = rcd_to_iport(rcd);579		__le32 *rhf_addr = (__le32 *)rcd->rcvhdrq + mdata.ps_head +580					 packet->rcd->rhf_offset;581		struct rvt_qp *qp;582		struct ib_header *hdr;583		struct rvt_dev_info *rdi = &rcd->dd->verbs_dev.rdi;584		u64 rhf = rhf_to_cpu(rhf_addr);585		u32 etype = rhf_rcv_type(rhf), qpn, bth1;586		u8 lnh;587 588		if (ps_done(&mdata, rhf, rcd))589			break;590 591		if (ps_skip(&mdata, rhf, rcd))592			goto next;593 594		if (etype != RHF_RCV_TYPE_IB)595			goto next;596 597		packet->hdr = hfi1_get_msgheader(packet->rcd, rhf_addr);598		hdr = packet->hdr;599		lnh = ib_get_lnh(hdr);600 601		if (lnh == HFI1_LRH_BTH) {602			packet->ohdr = &hdr->u.oth;603			packet->grh = NULL;604		} else if (lnh == HFI1_LRH_GRH) {605			packet->ohdr = &hdr->u.l.oth;606			packet->grh = &hdr->u.l.grh;607		} else {608			goto next; /* just in case */609		}610 611		if (!hfi1_may_ecn(packet))612			goto next;613 614		bth1 = be32_to_cpu(packet->ohdr->bth[1]);615		qpn = bth1 & RVT_QPN_MASK;616		rcu_read_lock();617		qp = rvt_lookup_qpn(rdi, &ibp->rvp, qpn);618 619		if (!qp) {620			rcu_read_unlock();621			goto next;622		}623 624		hfi1_process_ecn_slowpath(qp, packet, true);625		rcu_read_unlock();626 627		/* turn off BECN, FECN */628		bth1 &= ~(IB_FECN_SMASK | IB_BECN_SMASK);629		packet->ohdr->bth[1] = cpu_to_be32(bth1);630next:631		update_ps_mdata(&mdata, rcd);632	}633}634 635static void process_rcv_qp_work(struct hfi1_packet *packet)636{637	struct rvt_qp *qp, *nqp;638	struct hfi1_ctxtdata *rcd = packet->rcd;639 640	/*641	 * Iterate over all QPs waiting to respond.642	 * The list won't change since the IRQ is only run on one CPU.643	 */644	list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {645		list_del_init(&qp->rspwait);646		if (qp->r_flags & RVT_R_RSP_NAK) {647			qp->r_flags &= ~RVT_R_RSP_NAK;648			packet->qp = qp;649			hfi1_send_rc_ack(packet, 0);650		}651		if (qp->r_flags & RVT_R_RSP_SEND) {652			unsigned long flags;653 654			qp->r_flags &= ~RVT_R_RSP_SEND;655			spin_lock_irqsave(&qp->s_lock, flags);656			if (ib_rvt_state_ops[qp->state] &657					RVT_PROCESS_OR_FLUSH_SEND)658				hfi1_schedule_send(qp);659			spin_unlock_irqrestore(&qp->s_lock, flags);660		}661		rvt_put_qp(qp);662	}663}664 665static noinline int max_packet_exceeded(struct hfi1_packet *packet, int thread)666{667	if (thread) {668		if ((packet->numpkt & (MAX_PKT_RECV_THREAD - 1)) == 0)669			/* allow defered processing */670			process_rcv_qp_work(packet);671		cond_resched();672		return RCV_PKT_OK;673	} else {674		this_cpu_inc(*packet->rcd->dd->rcv_limit);675		return RCV_PKT_LIMIT;676	}677}678 679static inline int check_max_packet(struct hfi1_packet *packet, int thread)680{681	int ret = RCV_PKT_OK;682 683	if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0))684		ret = max_packet_exceeded(packet, thread);685	return ret;686}687 688static noinline int skip_rcv_packet(struct hfi1_packet *packet, int thread)689{690	int ret;691 692	packet->rcd->dd->ctx0_seq_drop++;693	/* Set up for the next packet */694	packet->rhqoff += packet->rsize;695	if (packet->rhqoff >= packet->maxcnt)696		packet->rhqoff = 0;697 698	packet->numpkt++;699	ret = check_max_packet(packet, thread);700 701	packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +702				     packet->rcd->rhf_offset;703	packet->rhf = rhf_to_cpu(packet->rhf_addr);704 705	return ret;706}707 708static void process_rcv_packet_napi(struct hfi1_packet *packet)709{710	packet->etype = rhf_rcv_type(packet->rhf);711 712	/* total length */713	packet->tlen = rhf_pkt_len(packet->rhf); /* in bytes */714	/* retrieve eager buffer details */715	packet->etail = rhf_egr_index(packet->rhf);716	packet->ebuf = get_egrbuf(packet->rcd, packet->rhf,717				  &packet->updegr);718	/*719	 * Prefetch the contents of the eager buffer.  It is720	 * OK to send a negative length to prefetch_range().721	 * The +2 is the size of the RHF.722	 */723	prefetch_range(packet->ebuf,724		       packet->tlen - ((packet->rcd->rcvhdrqentsize -725				       (rhf_hdrq_offset(packet->rhf)726					+ 2)) * 4));727 728	packet->rcd->rhf_rcv_function_map[packet->etype](packet);729	packet->numpkt++;730 731	/* Set up for the next packet */732	packet->rhqoff += packet->rsize;733	if (packet->rhqoff >= packet->maxcnt)734		packet->rhqoff = 0;735 736	packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +737				      packet->rcd->rhf_offset;738	packet->rhf = rhf_to_cpu(packet->rhf_addr);739}740 741static inline int process_rcv_packet(struct hfi1_packet *packet, int thread)742{743	int ret;744 745	packet->etype = rhf_rcv_type(packet->rhf);746 747	/* total length */748	packet->tlen = rhf_pkt_len(packet->rhf); /* in bytes */749	/* retrieve eager buffer details */750	packet->ebuf = NULL;751	if (rhf_use_egr_bfr(packet->rhf)) {752		packet->etail = rhf_egr_index(packet->rhf);753		packet->ebuf = get_egrbuf(packet->rcd, packet->rhf,754				 &packet->updegr);755		/*756		 * Prefetch the contents of the eager buffer.  It is757		 * OK to send a negative length to prefetch_range().758		 * The +2 is the size of the RHF.759		 */760		prefetch_range(packet->ebuf,761			       packet->tlen - ((get_hdrqentsize(packet->rcd) -762					       (rhf_hdrq_offset(packet->rhf)763						+ 2)) * 4));764	}765 766	/*767	 * Call a type specific handler for the packet. We768	 * should be able to trust that etype won't be beyond769	 * the range of valid indexes. If so something is really770	 * wrong and we can probably just let things come771	 * crashing down. There is no need to eat another772	 * comparison in this performance critical code.773	 */774	packet->rcd->rhf_rcv_function_map[packet->etype](packet);775	packet->numpkt++;776 777	/* Set up for the next packet */778	packet->rhqoff += packet->rsize;779	if (packet->rhqoff >= packet->maxcnt)780		packet->rhqoff = 0;781 782	ret = check_max_packet(packet, thread);783 784	packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +785				      packet->rcd->rhf_offset;786	packet->rhf = rhf_to_cpu(packet->rhf_addr);787 788	return ret;789}790 791static inline void process_rcv_update(int last, struct hfi1_packet *packet)792{793	/*794	 * Update head regs etc., every 16 packets, if not last pkt,795	 * to help prevent rcvhdrq overflows, when many packets796	 * are processed and queue is nearly full.797	 * Don't request an interrupt for intermediate updates.798	 */799	if (!last && !(packet->numpkt & 0xf)) {800		update_usrhead(packet->rcd, packet->rhqoff, packet->updegr,801			       packet->etail, 0, 0);802		packet->updegr = 0;803	}804	packet->grh = NULL;805}806 807static inline void finish_packet(struct hfi1_packet *packet)808{809	/*810	 * Nothing we need to free for the packet.811	 *812	 * The only thing we need to do is a final update and call for an813	 * interrupt814	 */815	update_usrhead(packet->rcd, hfi1_rcd_head(packet->rcd), packet->updegr,816		       packet->etail, rcv_intr_dynamic, packet->numpkt);817}818 819/*820 * handle_receive_interrupt_napi_fp - receive a packet821 * @rcd: the context822 * @budget: polling budget823 *824 * Called from interrupt handler for receive interrupt.825 * This is the fast path interrupt handler826 * when executing napi soft irq environment.827 */828int handle_receive_interrupt_napi_fp(struct hfi1_ctxtdata *rcd, int budget)829{830	struct hfi1_packet packet;831 832	init_packet(rcd, &packet);833	if (last_rcv_seq(rcd, rhf_rcv_seq(packet.rhf)))834		goto bail;835 836	while (packet.numpkt < budget) {837		process_rcv_packet_napi(&packet);838		if (hfi1_seq_incr(rcd, rhf_rcv_seq(packet.rhf)))839			break;840 841		process_rcv_update(0, &packet);842	}843	hfi1_set_rcd_head(rcd, packet.rhqoff);844bail:845	finish_packet(&packet);846	return packet.numpkt;847}848 849/*850 * Handle receive interrupts when using the no dma rtail option.851 */852int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata *rcd, int thread)853{854	int last = RCV_PKT_OK;855	struct hfi1_packet packet;856 857	init_packet(rcd, &packet);858	if (last_rcv_seq(rcd, rhf_rcv_seq(packet.rhf))) {859		last = RCV_PKT_DONE;860		goto bail;861	}862 863	prescan_rxq(rcd, &packet);864 865	while (last == RCV_PKT_OK) {866		last = process_rcv_packet(&packet, thread);867		if (hfi1_seq_incr(rcd, rhf_rcv_seq(packet.rhf)))868			last = RCV_PKT_DONE;869		process_rcv_update(last, &packet);870	}871	process_rcv_qp_work(&packet);872	hfi1_set_rcd_head(rcd, packet.rhqoff);873bail:874	finish_packet(&packet);875	return last;876}877 878int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata *rcd, int thread)879{880	u32 hdrqtail;881	int last = RCV_PKT_OK;882	struct hfi1_packet packet;883 884	init_packet(rcd, &packet);885	hdrqtail = get_rcvhdrtail(rcd);886	if (packet.rhqoff == hdrqtail) {887		last = RCV_PKT_DONE;888		goto bail;889	}890	smp_rmb();  /* prevent speculative reads of dma'ed hdrq */891 892	prescan_rxq(rcd, &packet);893 894	while (last == RCV_PKT_OK) {895		last = process_rcv_packet(&packet, thread);896		if (packet.rhqoff == hdrqtail)897			last = RCV_PKT_DONE;898		process_rcv_update(last, &packet);899	}900	process_rcv_qp_work(&packet);901	hfi1_set_rcd_head(rcd, packet.rhqoff);902bail:903	finish_packet(&packet);904	return last;905}906 907static void set_all_fastpath(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)908{909	u16 i;910 911	/*912	 * For dynamically allocated kernel contexts (like vnic) switch913	 * interrupt handler only for that context. Otherwise, switch914	 * interrupt handler for all statically allocated kernel contexts.915	 */916	if (rcd->ctxt >= dd->first_dyn_alloc_ctxt && !rcd->is_vnic) {917		hfi1_rcd_get(rcd);918		hfi1_set_fast(rcd);919		hfi1_rcd_put(rcd);920		return;921	}922 923	for (i = HFI1_CTRL_CTXT + 1; i < dd->num_rcv_contexts; i++) {924		rcd = hfi1_rcd_get_by_index(dd, i);925		if (rcd && (i < dd->first_dyn_alloc_ctxt || rcd->is_vnic))926			hfi1_set_fast(rcd);927		hfi1_rcd_put(rcd);928	}929}930 931void set_all_slowpath(struct hfi1_devdata *dd)932{933	struct hfi1_ctxtdata *rcd;934	u16 i;935 936	/* HFI1_CTRL_CTXT must always use the slow path interrupt handler */937	for (i = HFI1_CTRL_CTXT + 1; i < dd->num_rcv_contexts; i++) {938		rcd = hfi1_rcd_get_by_index(dd, i);939		if (!rcd)940			continue;941		if (i < dd->first_dyn_alloc_ctxt || rcd->is_vnic)942			rcd->do_interrupt = rcd->slow_handler;943 944		hfi1_rcd_put(rcd);945	}946}947 948static bool __set_armed_to_active(struct hfi1_packet *packet)949{950	u8 etype = rhf_rcv_type(packet->rhf);951	u8 sc = SC15_PACKET;952 953	if (etype == RHF_RCV_TYPE_IB) {954		struct ib_header *hdr = hfi1_get_msgheader(packet->rcd,955							   packet->rhf_addr);956		sc = hfi1_9B_get_sc5(hdr, packet->rhf);957	} else if (etype == RHF_RCV_TYPE_BYPASS) {958		struct hfi1_16b_header *hdr = hfi1_get_16B_header(959						packet->rcd,960						packet->rhf_addr);961		sc = hfi1_16B_get_sc(hdr);962	}963	if (sc != SC15_PACKET) {964		int hwstate = driver_lstate(packet->rcd->ppd);965		struct work_struct *lsaw =966				&packet->rcd->ppd->linkstate_active_work;967 968		if (hwstate != IB_PORT_ACTIVE) {969			dd_dev_info(packet->rcd->dd,970				    "Unexpected link state %s\n",971				    opa_lstate_name(hwstate));972			return false;973		}974 975		queue_work(packet->rcd->ppd->link_wq, lsaw);976		return true;977	}978	return false;979}980 981/**982 * set_armed_to_active  - the fast path for armed to active983 * @packet: the packet structure984 *985 * Return true if packet processing needs to bail.986 */987static bool set_armed_to_active(struct hfi1_packet *packet)988{989	if (likely(packet->rcd->ppd->host_link_state != HLS_UP_ARMED))990		return false;991	return __set_armed_to_active(packet);992}993 994/*995 * handle_receive_interrupt - receive a packet996 * @rcd: the context997 *998 * Called from interrupt handler for errors or receive interrupt.999 * This is the slow path interrupt handler.1000 */1001int handle_receive_interrupt(struct hfi1_ctxtdata *rcd, int thread)1002{1003	struct hfi1_devdata *dd = rcd->dd;1004	u32 hdrqtail;1005	int needset, last = RCV_PKT_OK;1006	struct hfi1_packet packet;1007	int skip_pkt = 0;1008 1009	if (!rcd->rcvhdrq)1010		return RCV_PKT_OK;1011	/* Control context will always use the slow path interrupt handler */1012	needset = (rcd->ctxt == HFI1_CTRL_CTXT) ? 0 : 1;1013 1014	init_packet(rcd, &packet);1015 1016	if (!get_dma_rtail_setting(rcd)) {1017		if (last_rcv_seq(rcd, rhf_rcv_seq(packet.rhf))) {1018			last = RCV_PKT_DONE;1019			goto bail;1020		}1021		hdrqtail = 0;1022	} else {1023		hdrqtail = get_rcvhdrtail(rcd);1024		if (packet.rhqoff == hdrqtail) {1025			last = RCV_PKT_DONE;1026			goto bail;1027		}1028		smp_rmb();  /* prevent speculative reads of dma'ed hdrq */1029 1030		/*1031		 * Control context can potentially receive an invalid1032		 * rhf. Drop such packets.1033		 */1034		if (rcd->ctxt == HFI1_CTRL_CTXT)1035			if (last_rcv_seq(rcd, rhf_rcv_seq(packet.rhf)))1036				skip_pkt = 1;1037	}1038 1039	prescan_rxq(rcd, &packet);1040 1041	while (last == RCV_PKT_OK) {1042		if (hfi1_need_drop(dd)) {1043			/* On to the next packet */1044			packet.rhqoff += packet.rsize;1045			packet.rhf_addr = (__le32 *)rcd->rcvhdrq +1046					  packet.rhqoff +1047					  rcd->rhf_offset;1048			packet.rhf = rhf_to_cpu(packet.rhf_addr);1049 1050		} else if (skip_pkt) {1051			last = skip_rcv_packet(&packet, thread);1052			skip_pkt = 0;1053		} else {1054			if (set_armed_to_active(&packet))1055				goto bail;1056			last = process_rcv_packet(&packet, thread);1057		}1058 1059		if (!get_dma_rtail_setting(rcd)) {1060			if (hfi1_seq_incr(rcd, rhf_rcv_seq(packet.rhf)))1061				last = RCV_PKT_DONE;1062		} else {1063			if (packet.rhqoff == hdrqtail)1064				last = RCV_PKT_DONE;1065			/*1066			 * Control context can potentially receive an invalid1067			 * rhf. Drop such packets.1068			 */1069			if (rcd->ctxt == HFI1_CTRL_CTXT) {1070				bool lseq;1071 1072				lseq = hfi1_seq_incr(rcd,1073						     rhf_rcv_seq(packet.rhf));1074				if (!last && lseq)1075					skip_pkt = 1;1076			}1077		}1078 1079		if (needset) {1080			needset = false;1081			set_all_fastpath(dd, rcd);1082		}1083		process_rcv_update(last, &packet);1084	}1085 1086	process_rcv_qp_work(&packet);1087	hfi1_set_rcd_head(rcd, packet.rhqoff);1088 1089bail:1090	/*1091	 * Always write head at end, and setup rcv interrupt, even1092	 * if no packets were processed.1093	 */1094	finish_packet(&packet);1095	return last;1096}1097 1098/*1099 * handle_receive_interrupt_napi_sp - receive a packet1100 * @rcd: the context1101 * @budget: polling budget1102 *1103 * Called from interrupt handler for errors or receive interrupt.1104 * This is the slow path interrupt handler1105 * when executing napi soft irq environment.1106 */1107int handle_receive_interrupt_napi_sp(struct hfi1_ctxtdata *rcd, int budget)1108{1109	struct hfi1_devdata *dd = rcd->dd;1110	int last = RCV_PKT_OK;1111	bool needset = true;1112	struct hfi1_packet packet;1113 1114	init_packet(rcd, &packet);1115	if (last_rcv_seq(rcd, rhf_rcv_seq(packet.rhf)))1116		goto bail;1117 1118	while (last != RCV_PKT_DONE && packet.numpkt < budget) {1119		if (hfi1_need_drop(dd)) {1120			/* On to the next packet */1121			packet.rhqoff += packet.rsize;1122			packet.rhf_addr = (__le32 *)rcd->rcvhdrq +1123					  packet.rhqoff +1124					  rcd->rhf_offset;1125			packet.rhf = rhf_to_cpu(packet.rhf_addr);1126 1127		} else {1128			if (set_armed_to_active(&packet))1129				goto bail;1130			process_rcv_packet_napi(&packet);1131		}1132 1133		if (hfi1_seq_incr(rcd, rhf_rcv_seq(packet.rhf)))1134			last = RCV_PKT_DONE;1135 1136		if (needset) {1137			needset = false;1138			set_all_fastpath(dd, rcd);1139		}1140 1141		process_rcv_update(last, &packet);1142	}1143 1144	hfi1_set_rcd_head(rcd, packet.rhqoff);1145 1146bail:1147	/*1148	 * Always write head at end, and setup rcv interrupt, even1149	 * if no packets were processed.1150	 */1151	finish_packet(&packet);1152	return packet.numpkt;1153}1154 1155/*1156 * We may discover in the interrupt that the hardware link state has1157 * changed from ARMED to ACTIVE (due to the arrival of a non-SC15 packet),1158 * and we need to update the driver's notion of the link state.  We cannot1159 * run set_link_state from interrupt context, so we queue this function on1160 * a workqueue.1161 *1162 * We delay the regular interrupt processing until after the state changes1163 * so that the link will be in the correct state by the time any application1164 * we wake up attempts to send a reply to any message it received.1165 * (Subsequent receive interrupts may possibly force the wakeup before we1166 * update the link state.)1167 *1168 * The rcd is freed in hfi1_free_ctxtdata after hfi1_postinit_cleanup invokes1169 * dd->f_cleanup(dd) to disable the interrupt handler and flush workqueues,1170 * so we're safe from use-after-free of the rcd.1171 */1172void receive_interrupt_work(struct work_struct *work)1173{1174	struct hfi1_pportdata *ppd = container_of(work, struct hfi1_pportdata,1175						  linkstate_active_work);1176	struct hfi1_devdata *dd = ppd->dd;1177	struct hfi1_ctxtdata *rcd;1178	u16 i;1179 1180	/* Received non-SC15 packet implies neighbor_normal */1181	ppd->neighbor_normal = 1;1182	set_link_state(ppd, HLS_UP_ACTIVE);1183 1184	/*1185	 * Interrupt all statically allocated kernel contexts that could1186	 * have had an interrupt during auto activation.1187	 */1188	for (i = HFI1_CTRL_CTXT; i < dd->first_dyn_alloc_ctxt; i++) {1189		rcd = hfi1_rcd_get_by_index(dd, i);1190		if (rcd)1191			force_recv_intr(rcd);1192		hfi1_rcd_put(rcd);1193	}1194}1195 1196/*1197 * Convert a given MTU size to the on-wire MAD packet enumeration.1198 * Return -1 if the size is invalid.1199 */1200int mtu_to_enum(u32 mtu, int default_if_bad)1201{1202	switch (mtu) {1203	case     0: return OPA_MTU_0;1204	case   256: return OPA_MTU_256;1205	case   512: return OPA_MTU_512;1206	case  1024: return OPA_MTU_1024;1207	case  2048: return OPA_MTU_2048;1208	case  4096: return OPA_MTU_4096;1209	case  8192: return OPA_MTU_8192;1210	case 10240: return OPA_MTU_10240;1211	}1212	return default_if_bad;1213}1214 1215u16 enum_to_mtu(int mtu)1216{1217	switch (mtu) {1218	case OPA_MTU_0:     return 0;1219	case OPA_MTU_256:   return 256;1220	case OPA_MTU_512:   return 512;1221	case OPA_MTU_1024:  return 1024;1222	case OPA_MTU_2048:  return 2048;1223	case OPA_MTU_4096:  return 4096;1224	case OPA_MTU_8192:  return 8192;1225	case OPA_MTU_10240: return 10240;1226	default: return 0xffff;1227	}1228}1229 1230/*1231 * set_mtu - set the MTU1232 * @ppd: the per port data1233 *1234 * We can handle "any" incoming size, the issue here is whether we1235 * need to restrict our outgoing size.  We do not deal with what happens1236 * to programs that are already running when the size changes.1237 */1238int set_mtu(struct hfi1_pportdata *ppd)1239{1240	struct hfi1_devdata *dd = ppd->dd;1241	int i, drain, ret = 0, is_up = 0;1242 1243	ppd->ibmtu = 0;1244	for (i = 0; i < ppd->vls_supported; i++)1245		if (ppd->ibmtu < dd->vld[i].mtu)1246			ppd->ibmtu = dd->vld[i].mtu;1247	ppd->ibmaxlen = ppd->ibmtu + lrh_max_header_bytes(ppd->dd);1248 1249	mutex_lock(&ppd->hls_lock);1250	if (ppd->host_link_state == HLS_UP_INIT ||1251	    ppd->host_link_state == HLS_UP_ARMED ||1252	    ppd->host_link_state == HLS_UP_ACTIVE)1253		is_up = 1;1254 1255	drain = !is_ax(dd) && is_up;1256 1257	if (drain)1258		/*1259		 * MTU is specified per-VL. To ensure that no packet gets1260		 * stuck (due, e.g., to the MTU for the packet's VL being1261		 * reduced), empty the per-VL FIFOs before adjusting MTU.1262		 */1263		ret = stop_drain_data_vls(dd);1264 1265	if (ret) {1266		dd_dev_err(dd, "%s: cannot stop/drain VLs - refusing to change per-VL MTUs\n",1267			   __func__);1268		goto err;1269	}1270 1271	hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_MTU, 0);1272 1273	if (drain)1274		open_fill_data_vls(dd); /* reopen all VLs */1275 1276err:1277	mutex_unlock(&ppd->hls_lock);1278 1279	return ret;1280}1281 1282int hfi1_set_lid(struct hfi1_pportdata *ppd, u32 lid, u8 lmc)1283{1284	struct hfi1_devdata *dd = ppd->dd;1285 1286	ppd->lid = lid;1287	ppd->lmc = lmc;1288	hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_LIDLMC, 0);1289 1290	dd_dev_info(dd, "port %u: got a lid: 0x%x\n", ppd->port, lid);1291 1292	return 0;1293}1294 1295void shutdown_led_override(struct hfi1_pportdata *ppd)1296{1297	struct hfi1_devdata *dd = ppd->dd;1298 1299	/*1300	 * This pairs with the memory barrier in hfi1_start_led_override to1301	 * ensure that we read the correct state of LED beaconing represented1302	 * by led_override_timer_active1303	 */1304	smp_rmb();1305	if (atomic_read(&ppd->led_override_timer_active)) {1306		del_timer_sync(&ppd->led_override_timer);1307		atomic_set(&ppd->led_override_timer_active, 0);1308		/* Ensure the atomic_set is visible to all CPUs */1309		smp_wmb();1310	}1311 1312	/* Hand control of the LED to the DC for normal operation */1313	write_csr(dd, DCC_CFG_LED_CNTRL, 0);1314}1315 1316static void run_led_override(struct timer_list *t)1317{1318	struct hfi1_pportdata *ppd = from_timer(ppd, t, led_override_timer);1319	struct hfi1_devdata *dd = ppd->dd;1320	unsigned long timeout;1321	int phase_idx;1322 1323	if (!(dd->flags & HFI1_INITTED))1324		return;1325 1326	phase_idx = ppd->led_override_phase & 1;1327 1328	setextled(dd, phase_idx);1329 1330	timeout = ppd->led_override_vals[phase_idx];1331 1332	/* Set up for next phase */1333	ppd->led_override_phase = !ppd->led_override_phase;1334 1335	mod_timer(&ppd->led_override_timer, jiffies + timeout);1336}1337 1338/*1339 * To have the LED blink in a particular pattern, provide timeon and timeoff1340 * in milliseconds.1341 * To turn off custom blinking and return to normal operation, use1342 * shutdown_led_override()1343 */1344void hfi1_start_led_override(struct hfi1_pportdata *ppd, unsigned int timeon,1345			     unsigned int timeoff)1346{1347	if (!(ppd->dd->flags & HFI1_INITTED))1348		return;1349 1350	/* Convert to jiffies for direct use in timer */1351	ppd->led_override_vals[0] = msecs_to_jiffies(timeoff);1352	ppd->led_override_vals[1] = msecs_to_jiffies(timeon);1353 1354	/* Arbitrarily start from LED on phase */1355	ppd->led_override_phase = 1;1356 1357	/*1358	 * If the timer has not already been started, do so. Use a "quick"1359	 * timeout so the handler will be called soon to look at our request.1360	 */1361	if (!timer_pending(&ppd->led_override_timer)) {1362		timer_setup(&ppd->led_override_timer, run_led_override, 0);1363		ppd->led_override_timer.expires = jiffies + 1;1364		add_timer(&ppd->led_override_timer);1365		atomic_set(&ppd->led_override_timer_active, 1);1366		/* Ensure the atomic_set is visible to all CPUs */1367		smp_wmb();1368	}1369}1370 1371/**1372 * hfi1_reset_device - reset the chip if possible1373 * @unit: the device to reset1374 *1375 * Whether or not reset is successful, we attempt to re-initialize the chip1376 * (that is, much like a driver unload/reload).  We clear the INITTED flag1377 * so that the various entry points will fail until we reinitialize.  For1378 * now, we only allow this if no user contexts are open that use chip resources1379 */1380int hfi1_reset_device(int unit)1381{1382	int ret;1383	struct hfi1_devdata *dd = hfi1_lookup(unit);1384	struct hfi1_pportdata *ppd;1385	int pidx;1386 1387	if (!dd) {1388		ret = -ENODEV;1389		goto bail;1390	}1391 1392	dd_dev_info(dd, "Reset on unit %u requested\n", unit);1393 1394	if (!dd->kregbase1 || !(dd->flags & HFI1_PRESENT)) {1395		dd_dev_info(dd,1396			    "Invalid unit number %u or not initialized or not present\n",1397			    unit);1398		ret = -ENXIO;1399		goto bail;1400	}1401 1402	/* If there are any user/vnic contexts, we cannot reset */1403	mutex_lock(&hfi1_mutex);1404	if (dd->rcd)1405		if (hfi1_stats.sps_ctxts) {1406			mutex_unlock(&hfi1_mutex);1407			ret = -EBUSY;1408			goto bail;1409		}1410	mutex_unlock(&hfi1_mutex);1411 1412	for (pidx = 0; pidx < dd->num_pports; ++pidx) {1413		ppd = dd->pport + pidx;1414 1415		shutdown_led_override(ppd);1416	}1417	if (dd->flags & HFI1_HAS_SEND_DMA)1418		sdma_exit(dd);1419 1420	hfi1_reset_cpu_counters(dd);1421 1422	ret = hfi1_init(dd, 1);1423 1424	if (ret)1425		dd_dev_err(dd,1426			   "Reinitialize unit %u after reset failed with %d\n",1427			   unit, ret);1428	else1429		dd_dev_info(dd, "Reinitialized unit %u after resetting\n",1430			    unit);1431 1432bail:1433	return ret;1434}1435 1436static inline void hfi1_setup_ib_header(struct hfi1_packet *packet)1437{1438	packet->hdr = (struct hfi1_ib_message_header *)1439			hfi1_get_msgheader(packet->rcd,1440					   packet->rhf_addr);1441	packet->hlen = (u8 *)packet->rhf_addr - (u8 *)packet->hdr;1442}1443 1444static int hfi1_bypass_ingress_pkt_check(struct hfi1_packet *packet)1445{1446	struct hfi1_pportdata *ppd = packet->rcd->ppd;1447 1448	/* slid and dlid cannot be 0 */1449	if ((!packet->slid) || (!packet->dlid))1450		return -EINVAL;1451 1452	/* Compare port lid with incoming packet dlid */1453	if ((!(hfi1_is_16B_mcast(packet->dlid))) &&1454	    (packet->dlid !=1455		opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B))) {1456		if ((packet->dlid & ~((1 << ppd->lmc) - 1)) != ppd->lid)1457			return -EINVAL;1458	}1459 1460	/* No multicast packets with SC15 */1461	if ((hfi1_is_16B_mcast(packet->dlid)) && (packet->sc == 0xF))1462		return -EINVAL;1463 1464	/* Packets with permissive DLID always on SC15 */1465	if ((packet->dlid == opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE),1466					 16B)) &&1467	    (packet->sc != 0xF))1468		return -EINVAL;1469 1470	return 0;1471}1472 1473static int hfi1_setup_9B_packet(struct hfi1_packet *packet)1474{1475	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);1476	struct ib_header *hdr;1477	u8 lnh;1478 1479	hfi1_setup_ib_header(packet);1480	hdr = packet->hdr;1481 1482	lnh = ib_get_lnh(hdr);1483	if (lnh == HFI1_LRH_BTH) {1484		packet->ohdr = &hdr->u.oth;1485		packet->grh = NULL;1486	} else if (lnh == HFI1_LRH_GRH) {1487		u32 vtf;1488 1489		packet->ohdr = &hdr->u.l.oth;1490		packet->grh = &hdr->u.l.grh;1491		if (packet->grh->next_hdr != IB_GRH_NEXT_HDR)1492			goto drop;1493		vtf = be32_to_cpu(packet->grh->version_tclass_flow);1494		if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)1495			goto drop;1496	} else {1497		goto drop;1498	}1499 1500	/* Query commonly used fields from packet header */1501	packet->payload = packet->ebuf;1502	packet->opcode = ib_bth_get_opcode(packet->ohdr);1503	packet->slid = ib_get_slid(hdr);1504	packet->dlid = ib_get_dlid(hdr);1505	if (unlikely((packet->dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) &&1506		     (packet->dlid != be16_to_cpu(IB_LID_PERMISSIVE))))1507		packet->dlid += opa_get_mcast_base(OPA_MCAST_NR) -1508				be16_to_cpu(IB_MULTICAST_LID_BASE);1509	packet->sl = ib_get_sl(hdr);1510	packet->sc = hfi1_9B_get_sc5(hdr, packet->rhf);1511	packet->pad = ib_bth_get_pad(packet->ohdr);1512	packet->extra_byte = 0;1513	packet->pkey = ib_bth_get_pkey(packet->ohdr);1514	packet->migrated = ib_bth_is_migration(packet->ohdr);1515 1516	return 0;1517drop:1518	ibp->rvp.n_pkt_drops++;1519	return -EINVAL;1520}1521 1522static int hfi1_setup_bypass_packet(struct hfi1_packet *packet)1523{1524	/*1525	 * Bypass packets have a different header/payload split1526	 * compared to an IB packet.1527	 * Current split is set such that 16 bytes of the actual1528	 * header is in the header buffer and the remining is in1529	 * the eager buffer. We chose 16 since hfi1 driver only1530	 * supports 16B bypass packets and we will be able to1531	 * receive the entire LRH with such a split.1532	 */1533 1534	struct hfi1_ctxtdata *rcd = packet->rcd;1535	struct hfi1_pportdata *ppd = rcd->ppd;1536	struct hfi1_ibport *ibp = &ppd->ibport_data;1537	u8 l4;1538 1539	packet->hdr = (struct hfi1_16b_header *)1540			hfi1_get_16B_header(packet->rcd,1541					    packet->rhf_addr);1542	l4 = hfi1_16B_get_l4(packet->hdr);1543	if (l4 == OPA_16B_L4_IB_LOCAL) {1544		packet->ohdr = packet->ebuf;1545		packet->grh = NULL;1546		packet->opcode = ib_bth_get_opcode(packet->ohdr);1547		packet->pad = hfi1_16B_bth_get_pad(packet->ohdr);1548		/* hdr_len_by_opcode already has an IB LRH factored in */1549		packet->hlen = hdr_len_by_opcode[packet->opcode] +1550			(LRH_16B_BYTES - LRH_9B_BYTES);1551		packet->migrated = opa_bth_is_migration(packet->ohdr);1552	} else if (l4 == OPA_16B_L4_IB_GLOBAL) {1553		u32 vtf;1554		u8 grh_len = sizeof(struct ib_grh);1555 1556		packet->ohdr = packet->ebuf + grh_len;1557		packet->grh = packet->ebuf;1558		packet->opcode = ib_bth_get_opcode(packet->ohdr);1559		packet->pad = hfi1_16B_bth_get_pad(packet->ohdr);1560		/* hdr_len_by_opcode already has an IB LRH factored in */1561		packet->hlen = hdr_len_by_opcode[packet->opcode] +1562			(LRH_16B_BYTES - LRH_9B_BYTES) + grh_len;1563		packet->migrated = opa_bth_is_migration(packet->ohdr);1564 1565		if (packet->grh->next_hdr != IB_GRH_NEXT_HDR)1566			goto drop;1567		vtf = be32_to_cpu(packet->grh->version_tclass_flow);1568		if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)1569			goto drop;1570	} else if (l4 == OPA_16B_L4_FM) {1571		packet->mgmt = packet->ebuf;1572		packet->ohdr = NULL;1573		packet->grh = NULL;1574		packet->opcode = IB_OPCODE_UD_SEND_ONLY;1575		packet->pad = OPA_16B_L4_FM_PAD;1576		packet->hlen = OPA_16B_L4_FM_HLEN;1577		packet->migrated = false;1578	} else {1579		goto drop;1580	}1581 1582	/* Query commonly used fields from packet header */1583	packet->payload = packet->ebuf + packet->hlen - LRH_16B_BYTES;1584	packet->slid = hfi1_16B_get_slid(packet->hdr);1585	packet->dlid = hfi1_16B_get_dlid(packet->hdr);1586	if (unlikely(hfi1_is_16B_mcast(packet->dlid)))1587		packet->dlid += opa_get_mcast_base(OPA_MCAST_NR) -1588				opa_get_lid(opa_get_mcast_base(OPA_MCAST_NR),1589					    16B);1590	packet->sc = hfi1_16B_get_sc(packet->hdr);1591	packet->sl = ibp->sc_to_sl[packet->sc];1592	packet->extra_byte = SIZE_OF_LT;1593	packet->pkey = hfi1_16B_get_pkey(packet->hdr);1594 1595	if (hfi1_bypass_ingress_pkt_check(packet))1596		goto drop;1597 1598	return 0;1599drop:1600	hfi1_cdbg(PKT, "%s: packet dropped", __func__);1601	ibp->rvp.n_pkt_drops++;1602	return -EINVAL;1603}1604 1605static void show_eflags_errs(struct hfi1_packet *packet)1606{1607	struct hfi1_ctxtdata *rcd = packet->rcd;1608	u32 rte = rhf_rcv_type_err(packet->rhf);1609 1610	dd_dev_err(rcd->dd,1611		   "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s] rte 0x%x\n",1612		   rcd->ctxt, packet->rhf,1613		   packet->rhf & RHF_K_HDR_LEN_ERR ? "k_hdr_len " : "",1614		   packet->rhf & RHF_DC_UNC_ERR ? "dc_unc " : "",1615		   packet->rhf & RHF_DC_ERR ? "dc " : "",1616		   packet->rhf & RHF_TID_ERR ? "tid " : "",1617		   packet->rhf & RHF_LEN_ERR ? "len " : "",1618		   packet->rhf & RHF_ECC_ERR ? "ecc " : "",1619		   packet->rhf & RHF_ICRC_ERR ? "icrc " : "",1620		   rte);1621}1622 1623void handle_eflags(struct hfi1_packet *packet)1624{1625	struct hfi1_ctxtdata *rcd = packet->rcd;1626 1627	rcv_hdrerr(rcd, rcd->ppd, packet);1628	if (rhf_err_flags(packet->rhf))1629		show_eflags_errs(packet);1630}1631 1632static void hfi1_ipoib_ib_rcv(struct hfi1_packet *packet)1633{1634	struct hfi1_ibport *ibp;1635	struct net_device *netdev;1636	struct hfi1_ctxtdata *rcd = packet->rcd;1637	struct napi_struct *napi = rcd->napi;1638	struct sk_buff *skb;1639	struct hfi1_netdev_rxq *rxq = container_of(napi,1640			struct hfi1_netdev_rxq, napi);1641	u32 extra_bytes;1642	u32 tlen, qpnum;1643	bool do_work, do_cnp;1644 1645	trace_hfi1_rcvhdr(packet);1646 1647	hfi1_setup_ib_header(packet);1648 1649	packet->ohdr = &((struct ib_header *)packet->hdr)->u.oth;1650	packet->grh = NULL;1651 1652	if (unlikely(rhf_err_flags(packet->rhf))) {1653		handle_eflags(packet);1654		return;1655	}1656 1657	qpnum = ib_bth_get_qpn(packet->ohdr);1658	netdev = hfi1_netdev_get_data(rcd->dd, qpnum);1659	if (!netdev)1660		goto drop_no_nd;1661 1662	trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf)));1663	trace_ctxt_rsm_hist(rcd->ctxt);1664 1665	/* handle congestion notifications */1666	do_work = hfi1_may_ecn(packet);1667	if (unlikely(do_work)) {1668		do_cnp = (packet->opcode != IB_OPCODE_CNP);1669		(void)hfi1_process_ecn_slowpath(hfi1_ipoib_priv(netdev)->qp,1670						 packet, do_cnp);1671	}1672 1673	/*1674	 * We have split point after last byte of DETH1675	 * lets strip padding and CRC and ICRC.1676	 * tlen is whole packet len so we need to1677	 * subtract header size as well.1678	 */1679	tlen = packet->tlen;1680	extra_bytes = ib_bth_get_pad(packet->ohdr) + (SIZE_OF_CRC << 2) +1681			packet->hlen;1682	if (unlikely(tlen < extra_bytes))1683		goto drop;1684 1685	tlen -= extra_bytes;1686 1687	skb = hfi1_ipoib_prepare_skb(rxq, tlen, packet->ebuf);1688	if (unlikely(!skb))1689		goto drop;1690 1691	dev_sw_netstats_rx_add(netdev, skb->len);1692 1693	skb->dev = netdev;1694	skb->pkt_type = PACKET_HOST;1695	netif_receive_skb(skb);1696 1697	return;1698 1699drop:1700	++netdev->stats.rx_dropped;1701drop_no_nd:1702	ibp = rcd_to_iport(packet->rcd);1703	++ibp->rvp.n_pkt_drops;1704}1705 1706/*1707 * The following functions are called by the interrupt handler. They are type1708 * specific handlers for each packet type.1709 */1710static void process_receive_ib(struct hfi1_packet *packet)1711{1712	if (hfi1_setup_9B_packet(packet))1713		return;1714 1715	if (unlikely(hfi1_dbg_should_fault_rx(packet)))1716		return;1717 1718	trace_hfi1_rcvhdr(packet);1719 1720	if (unlikely(rhf_err_flags(packet->rhf))) {1721		handle_eflags(packet);1722		return;1723	}1724 1725	hfi1_ib_rcv(packet);1726}1727 1728static void process_receive_bypass(struct hfi1_packet *packet)1729{1730	struct hfi1_devdata *dd = packet->rcd->dd;1731 1732	if (hfi1_setup_bypass_packet(packet))1733		return;1734 1735	trace_hfi1_rcvhdr(packet);1736 1737	if (unlikely(rhf_err_flags(packet->rhf))) {1738		handle_eflags(packet);1739		return;1740	}1741 1742	if (hfi1_16B_get_l2(packet->hdr) == 0x2) {1743		hfi1_16B_rcv(packet);1744	} else {1745		dd_dev_err(dd,1746			   "Bypass packets other than 16B are not supported in normal operation. Dropping\n");1747		incr_cntr64(&dd->sw_rcv_bypass_packet_errors);1748		if (!(dd->err_info_rcvport.status_and_code &1749		      OPA_EI_STATUS_SMASK)) {1750			u64 *flits = packet->ebuf;1751 1752			if (flits && !(packet->rhf & RHF_LEN_ERR)) {1753				dd->err_info_rcvport.packet_flit1 = flits[0];1754				dd->err_info_rcvport.packet_flit2 =1755					packet->tlen > sizeof(flits[0]) ?1756					flits[1] : 0;1757			}1758			dd->err_info_rcvport.status_and_code |=1759				(OPA_EI_STATUS_SMASK | BAD_L2_ERR);1760		}1761	}1762}1763 1764static void process_receive_error(struct hfi1_packet *packet)1765{1766	/* KHdrHCRCErr -- KDETH packet with a bad HCRC */1767	if (unlikely(1768		 hfi1_dbg_fault_suppress_err(&packet->rcd->dd->verbs_dev) &&1769		 (rhf_rcv_type_err(packet->rhf) == RHF_RCV_TYPE_ERROR ||1770		  packet->rhf & RHF_DC_ERR)))1771		return;1772 1773	hfi1_setup_ib_header(packet);1774	handle_eflags(packet);1775 1776	if (unlikely(rhf_err_flags(packet->rhf)))1777		dd_dev_err(packet->rcd->dd,1778			   "Unhandled error packet received. Dropping.\n");1779}1780 1781static void kdeth_process_expected(struct hfi1_packet *packet)1782{1783	hfi1_setup_9B_packet(packet);1784	if (unlikely(hfi1_dbg_should_fault_rx(packet)))1785		return;1786 1787	if (unlikely(rhf_err_flags(packet->rhf))) {1788		struct hfi1_ctxtdata *rcd = packet->rcd;1789 1790		if (hfi1_handle_kdeth_eflags(rcd, rcd->ppd, packet))1791			return;1792	}1793 1794	hfi1_kdeth_expected_rcv(packet);1795}1796 1797static void kdeth_process_eager(struct hfi1_packet *packet)1798{1799	hfi1_setup_9B_packet(packet);1800	if (unlikely(hfi1_dbg_should_fault_rx(packet)))1801		return;1802 1803	trace_hfi1_rcvhdr(packet);1804	if (unlikely(rhf_err_flags(packet->rhf))) {1805		struct hfi1_ctxtdata *rcd = packet->rcd;1806 1807		show_eflags_errs(packet);1808		if (hfi1_handle_kdeth_eflags(rcd, rcd->ppd, packet))1809			return;1810	}1811 1812	hfi1_kdeth_eager_rcv(packet);1813}1814 1815static void process_receive_invalid(struct hfi1_packet *packet)1816{1817	dd_dev_err(packet->rcd->dd, "Invalid packet type %d. Dropping\n",1818		   rhf_rcv_type(packet->rhf));1819}1820 1821#define HFI1_RCVHDR_DUMP_MAX	51822 1823void seqfile_dump_rcd(struct seq_file *s, struct hfi1_ctxtdata *rcd)1824{1825	struct hfi1_packet packet;1826	struct ps_mdata mdata;1827	int i;1828 1829	seq_printf(s, "Rcd %u: RcvHdr cnt %u entsize %u %s ctrl 0x%08llx status 0x%08llx, head %llu tail %llu  sw head %u\n",1830		   rcd->ctxt, get_hdrq_cnt(rcd), get_hdrqentsize(rcd),1831		   get_dma_rtail_setting(rcd) ?1832		   "dma_rtail" : "nodma_rtail",1833		   read_kctxt_csr(rcd->dd, rcd->ctxt, RCV_CTXT_CTRL),1834		   read_kctxt_csr(rcd->dd, rcd->ctxt, RCV_CTXT_STATUS),1835		   read_uctxt_csr(rcd->dd, rcd->ctxt, RCV_HDR_HEAD) &1836		   RCV_HDR_HEAD_HEAD_MASK,1837		   read_uctxt_csr(rcd->dd, rcd->ctxt, RCV_HDR_TAIL),1838		   rcd->head);1839 1840	init_packet(rcd, &packet);1841	init_ps_mdata(&mdata, &packet);1842 1843	for (i = 0; i < HFI1_RCVHDR_DUMP_MAX; i++) {1844		__le32 *rhf_addr = (__le32 *)rcd->rcvhdrq + mdata.ps_head +1845					 rcd->rhf_offset;1846		struct ib_header *hdr;1847		u64 rhf = rhf_to_cpu(rhf_addr);1848		u32 etype = rhf_rcv_type(rhf), qpn;1849		u8 opcode;1850		u32 psn;1851		u8 lnh;1852 1853		if (ps_done(&mdata, rhf, rcd))1854			break;1855 1856		if (ps_skip(&mdata, rhf, rcd))1857			goto next;1858 1859		if (etype > RHF_RCV_TYPE_IB)1860			goto next;1861 1862		packet.hdr = hfi1_get_msgheader(rcd, rhf_addr);1863		hdr = packet.hdr;1864 1865		lnh = be16_to_cpu(hdr->lrh[0]) & 3;1866 1867		if (lnh == HFI1_LRH_BTH)1868			packet.ohdr = &hdr->u.oth;1869		else if (lnh == HFI1_LRH_GRH)1870			packet.ohdr = &hdr->u.l.oth;1871		else1872			goto next; /* just in case */1873 1874		opcode = (be32_to_cpu(packet.ohdr->bth[0]) >> 24);1875		qpn = be32_to_cpu(packet.ohdr->bth[1]) & RVT_QPN_MASK;1876		psn = mask_psn(be32_to_cpu(packet.ohdr->bth[2]));1877 1878		seq_printf(s, "\tEnt %u: opcode 0x%x, qpn 0x%x, psn 0x%x\n",1879			   mdata.ps_head, opcode, qpn, psn);1880next:1881		update_ps_mdata(&mdata, rcd);1882	}1883}1884 1885const rhf_rcv_function_ptr normal_rhf_rcv_functions[] = {1886	[RHF_RCV_TYPE_EXPECTED] = kdeth_process_expected,1887	[RHF_RCV_TYPE_EAGER] = kdeth_process_eager,1888	[RHF_RCV_TYPE_IB] = process_receive_ib,1889	[RHF_RCV_TYPE_ERROR] = process_receive_error,1890	[RHF_RCV_TYPE_BYPASS] = process_receive_bypass,1891	[RHF_RCV_TYPE_INVALID5] = process_receive_invalid,1892	[RHF_RCV_TYPE_INVALID6] = process_receive_invalid,1893	[RHF_RCV_TYPE_INVALID7] = process_receive_invalid,1894};1895 1896const rhf_rcv_function_ptr netdev_rhf_rcv_functions[] = {1897	[RHF_RCV_TYPE_EXPECTED] = process_receive_invalid,1898	[RHF_RCV_TYPE_EAGER] = process_receive_invalid,1899	[RHF_RCV_TYPE_IB] = hfi1_ipoib_ib_rcv,1900	[RHF_RCV_TYPE_ERROR] = process_receive_error,1901	[RHF_RCV_TYPE_BYPASS] = hfi1_vnic_bypass_rcv,1902	[RHF_RCV_TYPE_INVALID5] = process_receive_invalid,1903	[RHF_RCV_TYPE_INVALID6] = process_receive_invalid,1904	[RHF_RCV_TYPE_INVALID7] = process_receive_invalid,1905};1906