1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2/*
3 * Copyright(c) 2015 - 2018 Intel Corporation.
4 */
5
6#include <linux/spinlock.h>
7
8#include "hfi.h"
9#include "mad.h"
10#include "qp.h"
11#include "verbs_txreq.h"
12#include "trace.h"
13
14static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
15{
16	return (gid->global.interface_id == id &&
17		(gid->global.subnet_prefix == gid_prefix ||
18		 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
19}
20
21/*
22 *
23 * This should be called with the QP r_lock held.
24 *
25 * The s_lock will be acquired around the hfi1_migrate_qp() call.
26 */
27int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_packet *packet)
28{
29	__be64 guid;
30	unsigned long flags;
31	struct rvt_qp *qp = packet->qp;
32	u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
33	u32 dlid = packet->dlid;
34	u32 slid = packet->slid;
35	u32 sl = packet->sl;
36	bool migrated = packet->migrated;
37	u16 pkey = packet->pkey;
38
39	if (qp->s_mig_state == IB_MIG_ARMED && migrated) {
40		if (!packet->grh) {
41			if ((rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
42			     IB_AH_GRH) &&
43			    (packet->etype != RHF_RCV_TYPE_BYPASS))
44				return 1;
45		} else {
46			const struct ib_global_route *grh;
47
48			if (!(rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
49			      IB_AH_GRH))
50				return 1;
51			grh = rdma_ah_read_grh(&qp->alt_ah_attr);
52			guid = get_sguid(ibp, grh->sgid_index);
53			if (!gid_ok(&packet->grh->dgid, ibp->rvp.gid_prefix,
54				    guid))
55				return 1;
56			if (!gid_ok(
57				&packet->grh->sgid,
58				grh->dgid.global.subnet_prefix,
59				grh->dgid.global.interface_id))
60				return 1;
61		}
62		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), pkey,
63					    sc5, slid))) {
64			hfi1_bad_pkey(ibp, pkey, sl, 0, qp->ibqp.qp_num,
65				      slid, dlid);
66			return 1;
67		}
68		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
69		if (slid != rdma_ah_get_dlid(&qp->alt_ah_attr) ||
70		    ppd_from_ibp(ibp)->port !=
71			rdma_ah_get_port_num(&qp->alt_ah_attr))
72			return 1;
73		spin_lock_irqsave(&qp->s_lock, flags);
74		hfi1_migrate_qp(qp);
75		spin_unlock_irqrestore(&qp->s_lock, flags);
76	} else {
77		if (!packet->grh) {
78			if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
79			     IB_AH_GRH) &&
80			    (packet->etype != RHF_RCV_TYPE_BYPASS))
81				return 1;
82		} else {
83			const struct ib_global_route *grh;
84
85			if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
86						   IB_AH_GRH))
87				return 1;
88			grh = rdma_ah_read_grh(&qp->remote_ah_attr);
89			guid = get_sguid(ibp, grh->sgid_index);
90			if (!gid_ok(&packet->grh->dgid, ibp->rvp.gid_prefix,
91				    guid))
92				return 1;
93			if (!gid_ok(
94			     &packet->grh->sgid,
95			     grh->dgid.global.subnet_prefix,
96			     grh->dgid.global.interface_id))
97				return 1;
98		}
99		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), pkey,
100					    sc5, slid))) {
101			hfi1_bad_pkey(ibp, pkey, sl, 0, qp->ibqp.qp_num,
102				      slid, dlid);
103			return 1;
104		}
105		/* Validate the SLID. See Ch. 9.6.1.5 */
106		if ((slid != rdma_ah_get_dlid(&qp->remote_ah_attr)) ||
107		    ppd_from_ibp(ibp)->port != qp->port_num)
108			return 1;
109		if (qp->s_mig_state == IB_MIG_REARM && !migrated)
110			qp->s_mig_state = IB_MIG_ARMED;
111	}
112
113	return 0;
114}
115
116/**
117 * hfi1_make_grh - construct a GRH header
118 * @ibp: a pointer to the IB port
119 * @hdr: a pointer to the GRH header being constructed
120 * @grh: the global route address to send to
121 * @hwords: size of header after grh being sent in dwords
122 * @nwords: the number of 32 bit words of data being sent
123 *
124 * Return the size of the header in 32 bit words.
125 */
126u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
127		  const struct ib_global_route *grh, u32 hwords, u32 nwords)
128{
129	hdr->version_tclass_flow =
130		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
131			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
132			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
133	hdr->paylen = cpu_to_be16((hwords + nwords) << 2);
134	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
135	hdr->next_hdr = IB_GRH_NEXT_HDR;
136	hdr->hop_limit = grh->hop_limit;
137	/* The SGID is 32-bit aligned. */
138	hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
139	hdr->sgid.global.interface_id =
140		grh->sgid_index < HFI1_GUIDS_PER_PORT ?
141		get_sguid(ibp, grh->sgid_index) :
142		get_sguid(ibp, HFI1_PORT_GUID_INDEX);
143	hdr->dgid = grh->dgid;
144
145	/* GRH header size in 32-bit words. */
146	return sizeof(struct ib_grh) / sizeof(u32);
147}
148
149#define BTH2_OFFSET (offsetof(struct hfi1_sdma_header, \
150			      hdr.ibh.u.oth.bth[2]) / 4)
151
152/**
153 * build_ahg - create ahg in s_ahg
154 * @qp: a pointer to QP
155 * @npsn: the next PSN for the request/response
156 *
157 * This routine handles the AHG by allocating an ahg entry and causing the
158 * copy of the first middle.
159 *
160 * Subsequent middles use the copied entry, editing the
161 * PSN with 1 or 2 edits.
162 */
163static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
164{
165	struct hfi1_qp_priv *priv = qp->priv;
166
167	if (unlikely(qp->s_flags & HFI1_S_AHG_CLEAR))
168		clear_ahg(qp);
169	if (!(qp->s_flags & HFI1_S_AHG_VALID)) {
170		/* first middle that needs copy  */
171		if (qp->s_ahgidx < 0)
172			qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
173		if (qp->s_ahgidx >= 0) {
174			qp->s_ahgpsn = npsn;
175			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
176			/* save to protect a change in another thread */
177			priv->s_ahg->ahgidx = qp->s_ahgidx;
178			qp->s_flags |= HFI1_S_AHG_VALID;
179		}
180	} else {
181		/* subsequent middle after valid */
182		if (qp->s_ahgidx >= 0) {
183			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_USE_AHG;
184			priv->s_ahg->ahgidx = qp->s_ahgidx;
185			priv->s_ahg->ahgcount++;
186			priv->s_ahg->ahgdesc[0] =
187				sdma_build_ahg_descriptor(
188					(__force u16)cpu_to_be16((u16)npsn),
189					BTH2_OFFSET,
190					16,
191					16);
192			if ((npsn & 0xffff0000) !=
193					(qp->s_ahgpsn & 0xffff0000)) {
194				priv->s_ahg->ahgcount++;
195				priv->s_ahg->ahgdesc[1] =
196					sdma_build_ahg_descriptor(
197						(__force u16)cpu_to_be16(
198							(u16)(npsn >> 16)),
199						BTH2_OFFSET,
200						0,
201						16);
202			}
203		}
204	}
205}
206
207static inline void hfi1_make_ruc_bth(struct rvt_qp *qp,
208				     struct ib_other_headers *ohdr,
209				     u32 bth0, u32 bth1, u32 bth2)
210{
211	ohdr->bth[0] = cpu_to_be32(bth0);
212	ohdr->bth[1] = cpu_to_be32(bth1);
213	ohdr->bth[2] = cpu_to_be32(bth2);
214}
215
216/**
217 * hfi1_make_ruc_header_16B - build a 16B header
218 * @qp: the queue pair
219 * @ohdr: a pointer to the destination header memory
220 * @bth0: bth0 passed in from the RC/UC builder
221 * @bth1: bth1 passed in from the RC/UC builder
222 * @bth2: bth2 passed in from the RC/UC builder
223 * @middle: non zero implies indicates ahg "could" be used
224 * @ps: the current packet state
225 *
226 * This routine may disarm ahg under these situations:
227 * - packet needs a GRH
228 * - BECN needed
229 * - migration state not IB_MIG_MIGRATED
230 */
231static inline void hfi1_make_ruc_header_16B(struct rvt_qp *qp,
232					    struct ib_other_headers *ohdr,
233					    u32 bth0, u32 bth1, u32 bth2,
234					    int middle,
235					    struct hfi1_pkt_state *ps)
236{
237	struct hfi1_qp_priv *priv = qp->priv;
238	struct hfi1_ibport *ibp = ps->ibp;
239	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
240	u32 slid;
241	u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
242	u8 l4 = OPA_16B_L4_IB_LOCAL;
243	u8 extra_bytes = hfi1_get_16b_padding(
244				(ps->s_txreq->hdr_dwords << 2),
245				ps->s_txreq->s_cur_size);
246	u32 nwords = SIZE_OF_CRC + ((ps->s_txreq->s_cur_size +
247				 extra_bytes + SIZE_OF_LT) >> 2);
248	bool becn = false;
249
250	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
251	    hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) {
252		struct ib_grh *grh;
253		struct ib_global_route *grd =
254			rdma_ah_retrieve_grh(&qp->remote_ah_attr);
255		/*
256		 * Ensure OPA GIDs are transformed to IB gids
257		 * before creating the GRH.
258		 */
259		if (grd->sgid_index == OPA_GID_INDEX)
260			grd->sgid_index = 0;
261		grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh;
262		l4 = OPA_16B_L4_IB_GLOBAL;
263		ps->s_txreq->hdr_dwords +=
264			hfi1_make_grh(ibp, grh, grd,
265				      ps->s_txreq->hdr_dwords - LRH_16B_DWORDS,
266				      nwords);
267		middle = 0;
268	}
269
270	if (qp->s_mig_state == IB_MIG_MIGRATED)
271		bth1 |= OPA_BTH_MIG_REQ;
272	else
273		middle = 0;
274
275	if (qp->s_flags & RVT_S_ECN) {
276		qp->s_flags &= ~RVT_S_ECN;
277		/* we recently received a FECN, so return a BECN */
278		becn = true;
279		middle = 0;
280	}
281	if (middle)
282		build_ahg(qp, bth2);
283	else
284		qp->s_flags &= ~HFI1_S_AHG_VALID;
285
286	bth0 |= pkey;
287	bth0 |= extra_bytes << 20;
288	hfi1_make_ruc_bth(qp, ohdr, bth0, bth1, bth2);
289
290	if (!ppd->lid)
291		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
292	else
293		slid = ppd->lid |
294			(rdma_ah_get_path_bits(&qp->remote_ah_attr) &
295			((1 << ppd->lmc) - 1));
296
297	hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah,
298			  slid,
299			  opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr),
300				      16B),
301			  (ps->s_txreq->hdr_dwords + nwords) >> 1,
302			  pkey, becn, 0, l4, priv->s_sc);
303}
304
305/**
306 * hfi1_make_ruc_header_9B - build a 9B header
307 * @qp: the queue pair
308 * @ohdr: a pointer to the destination header memory
309 * @bth0: bth0 passed in from the RC/UC builder
310 * @bth1: bth1 passed in from the RC/UC builder
311 * @bth2: bth2 passed in from the RC/UC builder
312 * @middle: non zero implies indicates ahg "could" be used
313 * @ps: the current packet state
314 *
315 * This routine may disarm ahg under these situations:
316 * - packet needs a GRH
317 * - BECN needed
318 * - migration state not IB_MIG_MIGRATED
319 */
320static inline void hfi1_make_ruc_header_9B(struct rvt_qp *qp,
321					   struct ib_other_headers *ohdr,
322					   u32 bth0, u32 bth1, u32 bth2,
323					   int middle,
324					   struct hfi1_pkt_state *ps)
325{
326	struct hfi1_qp_priv *priv = qp->priv;
327	struct hfi1_ibport *ibp = ps->ibp;
328	u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
329	u16 lrh0 = HFI1_LRH_BTH;
330	u8 extra_bytes = -ps->s_txreq->s_cur_size & 3;
331	u32 nwords = SIZE_OF_CRC + ((ps->s_txreq->s_cur_size +
332					 extra_bytes) >> 2);
333
334	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
335		struct ib_grh *grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh;
336
337		lrh0 = HFI1_LRH_GRH;
338		ps->s_txreq->hdr_dwords +=
339			hfi1_make_grh(ibp, grh,
340				      rdma_ah_read_grh(&qp->remote_ah_attr),
341				      ps->s_txreq->hdr_dwords - LRH_9B_DWORDS,
342				      nwords);
343		middle = 0;
344	}
345	lrh0 |= (priv->s_sc & 0xf) << 12 |
346		(rdma_ah_get_sl(&qp->remote_ah_attr) & 0xf) << 4;
347
348	if (qp->s_mig_state == IB_MIG_MIGRATED)
349		bth0 |= IB_BTH_MIG_REQ;
350	else
351		middle = 0;
352
353	if (qp->s_flags & RVT_S_ECN) {
354		qp->s_flags &= ~RVT_S_ECN;
355		/* we recently received a FECN, so return a BECN */
356		bth1 |= (IB_BECN_MASK << IB_BECN_SHIFT);
357		middle = 0;
358	}
359	if (middle)
360		build_ahg(qp, bth2);
361	else
362		qp->s_flags &= ~HFI1_S_AHG_VALID;
363
364	bth0 |= pkey;
365	bth0 |= extra_bytes << 20;
366	hfi1_make_ruc_bth(qp, ohdr, bth0, bth1, bth2);
367	hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh,
368			 lrh0,
369			 ps->s_txreq->hdr_dwords + nwords,
370			 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B),
371			 ppd_from_ibp(ibp)->lid |
372				rdma_ah_get_path_bits(&qp->remote_ah_attr));
373}
374
375typedef void (*hfi1_make_ruc_hdr)(struct rvt_qp *qp,
376				  struct ib_other_headers *ohdr,
377				  u32 bth0, u32 bth1, u32 bth2, int middle,
378				  struct hfi1_pkt_state *ps);
379
380/* We support only two types - 9B and 16B for now */
381static const hfi1_make_ruc_hdr hfi1_ruc_header_tbl[2] = {
382	[HFI1_PKT_TYPE_9B] = &hfi1_make_ruc_header_9B,
383	[HFI1_PKT_TYPE_16B] = &hfi1_make_ruc_header_16B
384};
385
386void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
387			  u32 bth0, u32 bth1, u32 bth2, int middle,
388			  struct hfi1_pkt_state *ps)
389{
390	struct hfi1_qp_priv *priv = qp->priv;
391
392	/*
393	 * reset s_ahg/AHG fields
394	 *
395	 * This insures that the ahgentry/ahgcount
396	 * are at a non-AHG default to protect
397	 * build_verbs_tx_desc() from using
398	 * an include ahgidx.
399	 *
400	 * build_ahg() will modify as appropriate
401	 * to use the AHG feature.
402	 */
403	priv->s_ahg->tx_flags = 0;
404	priv->s_ahg->ahgcount = 0;
405	priv->s_ahg->ahgidx = 0;
406
407	/* Make the appropriate header */
408	hfi1_ruc_header_tbl[priv->hdr_type](qp, ohdr, bth0, bth1, bth2, middle,
409					    ps);
410}
411
412/* when sending, force a reschedule every one of these periods */
413#define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
414
415/**
416 * hfi1_schedule_send_yield - test for a yield required for QP
417 * send engine
418 * @qp: a pointer to QP
419 * @ps: a pointer to a structure with commonly lookup values for
420 *      the send engine progress
421 * @tid: true if it is the tid leg
422 *
423 * This routine checks if the time slice for the QP has expired
424 * for RC QPs, if so an additional work entry is queued. At this
425 * point, other QPs have an opportunity to be scheduled. It
426 * returns true if a yield is required, otherwise, false
427 * is returned.
428 */
429bool hfi1_schedule_send_yield(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
430			      bool tid)
431{
432	ps->pkts_sent = true;
433
434	if (unlikely(time_after(jiffies, ps->timeout))) {
435		if (!ps->in_thread ||
436		    workqueue_congested(ps->cpu, ps->ppd->hfi1_wq)) {
437			spin_lock_irqsave(&qp->s_lock, ps->flags);
438			if (!tid) {
439				qp->s_flags &= ~RVT_S_BUSY;
440				hfi1_schedule_send(qp);
441			} else {
442				struct hfi1_qp_priv *priv = qp->priv;
443
444				if (priv->s_flags &
445				    HFI1_S_TID_BUSY_SET) {
446					qp->s_flags &= ~RVT_S_BUSY;
447					priv->s_flags &=
448						~(HFI1_S_TID_BUSY_SET |
449						  RVT_S_BUSY);
450				} else {
451					priv->s_flags &= ~RVT_S_BUSY;
452				}
453				hfi1_schedule_tid_send(qp);
454			}
455
456			spin_unlock_irqrestore(&qp->s_lock, ps->flags);
457			this_cpu_inc(*ps->ppd->dd->send_schedule);
458			trace_hfi1_rc_expired_time_slice(qp, true);
459			return true;
460		}
461
462		cond_resched();
463		this_cpu_inc(*ps->ppd->dd->send_schedule);
464		ps->timeout = jiffies + ps->timeout_int;
465	}
466
467	trace_hfi1_rc_expired_time_slice(qp, false);
468	return false;
469}
470
471void hfi1_do_send_from_rvt(struct rvt_qp *qp)
472{
473	hfi1_do_send(qp, false);
474}
475
476void _hfi1_do_send(struct work_struct *work)
477{
478	struct iowait_work *w = container_of(work, struct iowait_work, iowork);
479	struct rvt_qp *qp = iowait_to_qp(w->iow);
480
481	hfi1_do_send(qp, true);
482}
483
484/**
485 * hfi1_do_send - perform a send on a QP
486 * @qp: a pointer to the QP
487 * @in_thread: true if in a workqueue thread
488 *
489 * Process entries in the send work queue until credit or queue is
490 * exhausted.  Only allow one CPU to send a packet per QP.
491 * Otherwise, two threads could send packets out of order.
492 */
493void hfi1_do_send(struct rvt_qp *qp, bool in_thread)
494{
495	struct hfi1_pkt_state ps;
496	struct hfi1_qp_priv *priv = qp->priv;
497	int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
498
499	ps.dev = to_idev(qp->ibqp.device);
500	ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
501	ps.ppd = ppd_from_ibp(ps.ibp);
502	ps.in_thread = in_thread;
503	ps.wait = iowait_get_ib_work(&priv->s_iowait);
504
505	trace_hfi1_rc_do_send(qp, in_thread);
506
507	switch (qp->ibqp.qp_type) {
508	case IB_QPT_RC:
509		if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
510				   ~((1 << ps.ppd->lmc) - 1)) ==
511				  ps.ppd->lid)) {
512			rvt_ruc_loopback(qp);
513			return;
514		}
515		make_req = hfi1_make_rc_req;
516		ps.timeout_int = qp->timeout_jiffies;
517		break;
518	case IB_QPT_UC:
519		if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
520				   ~((1 << ps.ppd->lmc) - 1)) ==
521				  ps.ppd->lid)) {
522			rvt_ruc_loopback(qp);
523			return;
524		}
525		make_req = hfi1_make_uc_req;
526		ps.timeout_int = SEND_RESCHED_TIMEOUT;
527		break;
528	default:
529		make_req = hfi1_make_ud_req;
530		ps.timeout_int = SEND_RESCHED_TIMEOUT;
531	}
532
533	spin_lock_irqsave(&qp->s_lock, ps.flags);
534
535	/* Return if we are already busy processing a work request. */
536	if (!hfi1_send_ok(qp)) {
537		if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
538			iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
539		spin_unlock_irqrestore(&qp->s_lock, ps.flags);
540		return;
541	}
542
543	qp->s_flags |= RVT_S_BUSY;
544
545	ps.timeout_int = ps.timeout_int / 8;
546	ps.timeout = jiffies + ps.timeout_int;
547	ps.cpu = priv->s_sde ? priv->s_sde->cpu :
548			cpumask_first(cpumask_of_node(ps.ppd->dd->node));
549	ps.pkts_sent = false;
550
551	/* insure a pre-built packet is handled  */
552	ps.s_txreq = get_waiting_verbs_txreq(ps.wait);
553	do {
554		/* Check for a constructed packet to be sent. */
555		if (ps.s_txreq) {
556			if (priv->s_flags & HFI1_S_TID_BUSY_SET)
557				qp->s_flags |= RVT_S_BUSY;
558			spin_unlock_irqrestore(&qp->s_lock, ps.flags);
559			/*
560			 * If the packet cannot be sent now, return and
561			 * the send engine will be woken up later.
562			 */
563			if (hfi1_verbs_send(qp, &ps))
564				return;
565
566			/* allow other tasks to run */
567			if (hfi1_schedule_send_yield(qp, &ps, false))
568				return;
569
570			spin_lock_irqsave(&qp->s_lock, ps.flags);
571		}
572	} while (make_req(qp, &ps));
573	iowait_starve_clear(ps.pkts_sent, &priv->s_iowait);
574	spin_unlock_irqrestore(&qp->s_lock, ps.flags);
575}
576