Linux Audio

Check our new training course

Loading...
v6.8
  1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
  2/*
  3 * Copyright(c) 2015 - 2020 Intel Corporation.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  4 */
  5
  6#include <linux/err.h>
  7#include <linux/vmalloc.h>
  8#include <linux/hash.h>
  9#include <linux/module.h>
 10#include <linux/seq_file.h>
 11#include <rdma/rdma_vt.h>
 12#include <rdma/rdmavt_qp.h>
 13#include <rdma/ib_verbs.h>
 14
 15#include "hfi.h"
 16#include "qp.h"
 17#include "trace.h"
 18#include "verbs_txreq.h"
 19
 20unsigned int hfi1_qp_table_size = 256;
 21module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO);
 22MODULE_PARM_DESC(qp_table_size, "QP table size");
 23
 24static void flush_tx_list(struct rvt_qp *qp);
 25static int iowait_sleep(
 26	struct sdma_engine *sde,
 27	struct iowait_work *wait,
 28	struct sdma_txreq *stx,
 29	unsigned int seq,
 30	bool pkts_sent);
 31static void iowait_wakeup(struct iowait *wait, int reason);
 32static void iowait_sdma_drained(struct iowait *wait);
 33static void qp_pio_drain(struct rvt_qp *qp);
 34
 35const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
 36[IB_WR_RDMA_WRITE] = {
 37	.length = sizeof(struct ib_rdma_wr),
 38	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 39},
 40
 41[IB_WR_RDMA_READ] = {
 42	.length = sizeof(struct ib_rdma_wr),
 43	.qpt_support = BIT(IB_QPT_RC),
 44	.flags = RVT_OPERATION_ATOMIC,
 45},
 46
 47[IB_WR_ATOMIC_CMP_AND_SWP] = {
 48	.length = sizeof(struct ib_atomic_wr),
 49	.qpt_support = BIT(IB_QPT_RC),
 50	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
 51},
 52
 53[IB_WR_ATOMIC_FETCH_AND_ADD] = {
 54	.length = sizeof(struct ib_atomic_wr),
 55	.qpt_support = BIT(IB_QPT_RC),
 56	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
 57},
 58
 59[IB_WR_RDMA_WRITE_WITH_IMM] = {
 60	.length = sizeof(struct ib_rdma_wr),
 61	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 62},
 63
 64[IB_WR_SEND] = {
 65	.length = sizeof(struct ib_send_wr),
 66	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
 67		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 68},
 69
 70[IB_WR_SEND_WITH_IMM] = {
 71	.length = sizeof(struct ib_send_wr),
 72	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
 73		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 74},
 75
 76[IB_WR_REG_MR] = {
 77	.length = sizeof(struct ib_reg_wr),
 78	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 79	.flags = RVT_OPERATION_LOCAL,
 80},
 81
 82[IB_WR_LOCAL_INV] = {
 83	.length = sizeof(struct ib_send_wr),
 84	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 85	.flags = RVT_OPERATION_LOCAL,
 86},
 87
 88[IB_WR_SEND_WITH_INV] = {
 89	.length = sizeof(struct ib_send_wr),
 90	.qpt_support = BIT(IB_QPT_RC),
 91},
 92
 93[IB_WR_OPFN] = {
 94	.length = sizeof(struct ib_atomic_wr),
 95	.qpt_support = BIT(IB_QPT_RC),
 96	.flags = RVT_OPERATION_USE_RESERVE,
 97},
 98
 99[IB_WR_TID_RDMA_WRITE] = {
100	.length = sizeof(struct ib_rdma_wr),
101	.qpt_support = BIT(IB_QPT_RC),
102	.flags = RVT_OPERATION_IGN_RNR_CNT,
103},
104
105};
106
107static void flush_list_head(struct list_head *l)
108{
109	while (!list_empty(l)) {
110		struct sdma_txreq *tx;
111
112		tx = list_first_entry(
113			l,
114			struct sdma_txreq,
115			list);
116		list_del_init(&tx->list);
117		hfi1_put_txreq(
118			container_of(tx, struct verbs_txreq, txreq));
119	}
120}
121
122static void flush_tx_list(struct rvt_qp *qp)
123{
124	struct hfi1_qp_priv *priv = qp->priv;
125
126	flush_list_head(&iowait_get_ib_work(&priv->s_iowait)->tx_head);
127	flush_list_head(&iowait_get_tid_work(&priv->s_iowait)->tx_head);
128}
129
130static void flush_iowait(struct rvt_qp *qp)
131{
132	struct hfi1_qp_priv *priv = qp->priv;
133	unsigned long flags;
134	seqlock_t *lock = priv->s_iowait.lock;
135
136	if (!lock)
137		return;
138	write_seqlock_irqsave(lock, flags);
139	if (!list_empty(&priv->s_iowait.list)) {
140		list_del_init(&priv->s_iowait.list);
141		priv->s_iowait.lock = NULL;
142		rvt_put_qp(qp);
143	}
144	write_sequnlock_irqrestore(lock, flags);
145}
146
147/*
 
 
 
 
 
 
 
 
 
148 * This function is what we would push to the core layer if we wanted to be a
149 * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
150 * to blindly pass the MTU enum value from the PathRecord to us.
151 */
152static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
153{
 
 
154	/* Constraining 10KB packets to 8KB packets */
155	if (mtu == (enum ib_mtu)OPA_MTU_10240)
156		mtu = (enum ib_mtu)OPA_MTU_8192;
157	return opa_mtu_enum_to_int((enum opa_mtu)mtu);
 
 
 
158}
159
160int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
161			 int attr_mask, struct ib_udata *udata)
162{
163	struct ib_qp *ibqp = &qp->ibqp;
164	struct hfi1_ibdev *dev = to_idev(ibqp->device);
165	struct hfi1_devdata *dd = dd_from_dev(dev);
166	u8 sc;
167
168	if (attr_mask & IB_QP_AV) {
169		sc = ah_to_sc(ibqp->device, &attr->ah_attr);
170		if (sc == 0xf)
171			return -EINVAL;
172
173		if (!qp_to_sdma_engine(qp, sc) &&
174		    dd->flags & HFI1_HAS_SEND_DMA)
175			return -EINVAL;
176
177		if (!qp_to_send_context(qp, sc))
178			return -EINVAL;
179	}
180
181	if (attr_mask & IB_QP_ALT_PATH) {
182		sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
183		if (sc == 0xf)
184			return -EINVAL;
185
186		if (!qp_to_sdma_engine(qp, sc) &&
187		    dd->flags & HFI1_HAS_SEND_DMA)
188			return -EINVAL;
189
190		if (!qp_to_send_context(qp, sc))
191			return -EINVAL;
192	}
193
194	return 0;
195}
196
197/*
198 * qp_set_16b - Set the hdr_type based on whether the slid or the
199 * dlid in the connection is extended. Only applicable for RC and UC
200 * QPs. UD QPs determine this on the fly from the ah in the wqe
201 */
202static inline void qp_set_16b(struct rvt_qp *qp)
203{
204	struct hfi1_pportdata *ppd;
205	struct hfi1_ibport *ibp;
206	struct hfi1_qp_priv *priv = qp->priv;
207
208	/* Update ah_attr to account for extended LIDs */
209	hfi1_update_ah_attr(qp->ibqp.device, &qp->remote_ah_attr);
210
211	/* Create 32 bit LIDs */
212	hfi1_make_opa_lid(&qp->remote_ah_attr);
213
214	if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH))
215		return;
216
217	ibp = to_iport(qp->ibqp.device, qp->port_num);
218	ppd = ppd_from_ibp(ibp);
219	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, &qp->remote_ah_attr);
220}
221
222void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
223		    int attr_mask, struct ib_udata *udata)
224{
225	struct ib_qp *ibqp = &qp->ibqp;
226	struct hfi1_qp_priv *priv = qp->priv;
227
228	if (attr_mask & IB_QP_AV) {
229		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
230		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
231		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
232		qp_set_16b(qp);
233	}
234
235	if (attr_mask & IB_QP_PATH_MIG_STATE &&
236	    attr->path_mig_state == IB_MIG_MIGRATED &&
237	    qp->s_mig_state == IB_MIG_ARMED) {
238		qp->s_flags |= HFI1_S_AHG_CLEAR;
239		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
240		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
241		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
242		qp_set_16b(qp);
243	}
244
245	opfn_qp_init(qp, attr, attr_mask);
246}
247
248/**
249 * hfi1_setup_wqe - set up the wqe
250 * @qp: The qp
251 * @wqe: The built wqe
252 * @call_send: Determine if the send should be posted or scheduled.
253 *
254 * Perform setup of the wqe.  This is called
255 * prior to inserting the wqe into the ring but after
256 * the wqe has been setup by RDMAVT. This function
257 * allows the driver the opportunity to perform
258 * validation and additional setup of the wqe.
259 *
260 * Returns 0 on success, -EINVAL on failure
261 *
262 */
263int hfi1_setup_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe, bool *call_send)
264{
265	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
266	struct rvt_ah *ah;
267	struct hfi1_pportdata *ppd;
268	struct hfi1_devdata *dd;
269
270	switch (qp->ibqp.qp_type) {
271	case IB_QPT_RC:
272		hfi1_setup_tid_rdma_wqe(qp, wqe);
273		fallthrough;
274	case IB_QPT_UC:
275		if (wqe->length > 0x80000000U)
276			return -EINVAL;
277		if (wqe->length > qp->pmtu)
278			*call_send = false;
279		break;
280	case IB_QPT_SMI:
281		/*
282		 * SM packets should exclusively use VL15 and their SL is
283		 * ignored (IBTA v1.3, Section 3.5.8.2). Therefore, when ah
284		 * is created, SL is 0 in most cases and as a result some
285		 * fields (vl and pmtu) in ah may not be set correctly,
286		 * depending on the SL2SC and SC2VL tables at the time.
287		 */
288		ppd = ppd_from_ibp(ibp);
289		dd = dd_from_ppd(ppd);
290		if (wqe->length > dd->vld[15].mtu)
291			return -EINVAL;
292		break;
293	case IB_QPT_GSI:
294	case IB_QPT_UD:
295		ah = rvt_get_swqe_ah(wqe);
296		if (wqe->length > (1 << ah->log_pmtu))
297			return -EINVAL;
298		if (ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)] == 0xf)
299			return -EINVAL;
300		break;
301	default:
302		break;
303	}
304
305	/*
306	 * System latency between send and schedule is large enough that
307	 * forcing call_send to true for piothreshold packets is necessary.
308	 */
309	if (wqe->length <= piothreshold)
310		*call_send = true;
311	return 0;
312}
313
314/**
315 * _hfi1_schedule_send - schedule progress
316 * @qp: the QP
317 *
318 * This schedules qp progress w/o regard to the s_flags.
319 *
320 * It is only used in the post send, which doesn't hold
321 * the s_lock.
322 */
323bool _hfi1_schedule_send(struct rvt_qp *qp)
324{
325	struct hfi1_qp_priv *priv = qp->priv;
326	struct hfi1_ibport *ibp =
327		to_iport(qp->ibqp.device, qp->port_num);
328	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
329	struct hfi1_devdata *dd = ppd->dd;
330
331	if (dd->flags & HFI1_SHUTDOWN)
332		return true;
333
334	return iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
335			       priv->s_sde ?
336			       priv->s_sde->cpu :
337			       cpumask_first(cpumask_of_node(dd->node)));
338}
339
340static void qp_pio_drain(struct rvt_qp *qp)
341{
342	struct hfi1_qp_priv *priv = qp->priv;
343
344	if (!priv->s_sendcontext)
345		return;
346	while (iowait_pio_pending(&priv->s_iowait)) {
347		write_seqlock_irq(&priv->s_sendcontext->waitlock);
348		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
349		write_sequnlock_irq(&priv->s_sendcontext->waitlock);
350		iowait_pio_drain(&priv->s_iowait);
351		write_seqlock_irq(&priv->s_sendcontext->waitlock);
352		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
353		write_sequnlock_irq(&priv->s_sendcontext->waitlock);
354	}
355}
356
357/**
358 * hfi1_schedule_send - schedule progress
359 * @qp: the QP
360 *
361 * This schedules qp progress and caller should hold
362 * the s_lock.
363 * @return true if the first leg is scheduled;
364 * false if the first leg is not scheduled.
365 */
366bool hfi1_schedule_send(struct rvt_qp *qp)
367{
368	lockdep_assert_held(&qp->s_lock);
369	if (hfi1_send_ok(qp)) {
370		_hfi1_schedule_send(qp);
371		return true;
372	}
373	if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
374		iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
375				IOWAIT_PENDING_IB);
376	return false;
377}
378
379static void hfi1_qp_schedule(struct rvt_qp *qp)
380{
381	struct hfi1_qp_priv *priv = qp->priv;
382	bool ret;
383
384	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_IB)) {
385		ret = hfi1_schedule_send(qp);
386		if (ret)
387			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
388	}
389	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_TID)) {
390		ret = hfi1_schedule_tid_send(qp);
391		if (ret)
392			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
393	}
394}
395
396void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
397{
398	unsigned long flags;
399
400	spin_lock_irqsave(&qp->s_lock, flags);
401	if (qp->s_flags & flag) {
402		qp->s_flags &= ~flag;
403		trace_hfi1_qpwakeup(qp, flag);
404		hfi1_qp_schedule(qp);
405	}
406	spin_unlock_irqrestore(&qp->s_lock, flags);
407	/* Notify hfi1_destroy_qp() if it is waiting. */
408	rvt_put_qp(qp);
409}
410
411void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
412{
413	struct hfi1_qp_priv *priv = qp->priv;
414
415	if (iowait_set_work_flag(wait) == IOWAIT_IB_SE) {
416		qp->s_flags &= ~RVT_S_BUSY;
417		/*
418		 * If we are sending a first-leg packet from the second leg,
419		 * we need to clear the busy flag from priv->s_flags to
420		 * avoid a race condition when the qp wakes up before
421		 * the call to hfi1_verbs_send() returns to the second
422		 * leg. In that case, the second leg will terminate without
423		 * being re-scheduled, resulting in failure to send TID RDMA
424		 * WRITE DATA and TID RDMA ACK packets.
425		 */
426		if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
427			priv->s_flags &= ~(HFI1_S_TID_BUSY_SET |
428					   RVT_S_BUSY);
429			iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
430		}
431	} else {
432		priv->s_flags &= ~RVT_S_BUSY;
433	}
434}
435
436static int iowait_sleep(
437	struct sdma_engine *sde,
438	struct iowait_work *wait,
439	struct sdma_txreq *stx,
440	uint seq,
441	bool pkts_sent)
442{
443	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
444	struct rvt_qp *qp;
445	struct hfi1_qp_priv *priv;
446	unsigned long flags;
447	int ret = 0;
448
449	qp = tx->qp;
450	priv = qp->priv;
451
452	spin_lock_irqsave(&qp->s_lock, flags);
453	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
454		/*
455		 * If we couldn't queue the DMA request, save the info
456		 * and try again later rather than destroying the
457		 * buffer and undoing the side effects of the copy.
458		 */
459		/* Make a common routine? */
460		list_add_tail(&stx->list, &wait->tx_head);
461		write_seqlock(&sde->waitlock);
462		if (sdma_progress(sde, seq, stx))
463			goto eagain;
464		if (list_empty(&priv->s_iowait.list)) {
465			struct hfi1_ibport *ibp =
466				to_iport(qp->ibqp.device, qp->port_num);
467
468			ibp->rvp.n_dmawait++;
469			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
470			iowait_get_priority(&priv->s_iowait);
471			iowait_queue(pkts_sent, &priv->s_iowait,
472				     &sde->dmawait);
473			priv->s_iowait.lock = &sde->waitlock;
474			trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
475			rvt_get_qp(qp);
476		}
477		write_sequnlock(&sde->waitlock);
478		hfi1_qp_unbusy(qp, wait);
479		spin_unlock_irqrestore(&qp->s_lock, flags);
480		ret = -EBUSY;
481	} else {
482		spin_unlock_irqrestore(&qp->s_lock, flags);
483		hfi1_put_txreq(tx);
484	}
485	return ret;
486eagain:
487	write_sequnlock(&sde->waitlock);
488	spin_unlock_irqrestore(&qp->s_lock, flags);
489	list_del_init(&stx->list);
490	return -EAGAIN;
491}
492
493static void iowait_wakeup(struct iowait *wait, int reason)
494{
495	struct rvt_qp *qp = iowait_to_qp(wait);
496
497	WARN_ON(reason != SDMA_AVAIL_REASON);
498	hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
499}
500
501static void iowait_sdma_drained(struct iowait *wait)
502{
503	struct rvt_qp *qp = iowait_to_qp(wait);
504	unsigned long flags;
505
506	/*
507	 * This happens when the send engine notes
508	 * a QP in the error state and cannot
509	 * do the flush work until that QP's
510	 * sdma work has finished.
511	 */
512	spin_lock_irqsave(&qp->s_lock, flags);
513	if (qp->s_flags & RVT_S_WAIT_DMA) {
514		qp->s_flags &= ~RVT_S_WAIT_DMA;
515		hfi1_schedule_send(qp);
516	}
517	spin_unlock_irqrestore(&qp->s_lock, flags);
518}
519
520static void hfi1_init_priority(struct iowait *w)
521{
522	struct rvt_qp *qp = iowait_to_qp(w);
523	struct hfi1_qp_priv *priv = qp->priv;
524
525	if (qp->s_flags & RVT_S_ACK_PENDING)
526		w->priority++;
527	if (priv->s_flags & RVT_S_ACK_PENDING)
528		w->priority++;
529}
530
531/**
532 * qp_to_sdma_engine - map a qp to a send engine
533 * @qp: the QP
534 * @sc5: the 5 bit sc
535 *
536 * Return:
537 * A send engine for the qp or NULL for SMI type qp.
538 */
539struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
540{
541	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
542	struct sdma_engine *sde;
543
544	if (!(dd->flags & HFI1_HAS_SEND_DMA))
545		return NULL;
546	switch (qp->ibqp.qp_type) {
547	case IB_QPT_SMI:
548		return NULL;
549	default:
550		break;
551	}
552	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
553	return sde;
554}
555
556/**
557 * qp_to_send_context - map a qp to a send context
558 * @qp: the QP
559 * @sc5: the 5 bit sc
560 *
561 * Return:
562 * A send context for the qp
563 */
564struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
565{
566	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
567
568	switch (qp->ibqp.qp_type) {
569	case IB_QPT_SMI:
570		/* SMA packets to VL15 */
571		return dd->vld[15].sc;
572	default:
573		break;
574	}
575
576	return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
577					  sc5);
578}
579
580static const char * const qp_type_str[] = {
581	"SMI", "GSI", "RC", "UC", "UD",
582};
583
584static int qp_idle(struct rvt_qp *qp)
585{
586	return
587		qp->s_last == qp->s_acked &&
588		qp->s_acked == qp->s_cur &&
589		qp->s_cur == qp->s_tail &&
590		qp->s_tail == qp->s_head;
591}
592
593/**
594 * qp_iter_print - print the qp information to seq_file
595 * @s: the seq_file to emit the qp information on
596 * @iter: the iterator for the qp hash list
597 */
598void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
599{
600	struct rvt_swqe *wqe;
601	struct rvt_qp *qp = iter->qp;
602	struct hfi1_qp_priv *priv = qp->priv;
603	struct sdma_engine *sde;
604	struct send_context *send_context;
605	struct rvt_ack_entry *e = NULL;
606	struct rvt_srq *srq = qp->ibqp.srq ?
607		ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
608
609	sde = qp_to_sdma_engine(qp, priv->s_sc);
610	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
611	send_context = qp_to_send_context(qp, priv->s_sc);
612	if (qp->s_ack_queue)
613		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
614	seq_printf(s,
615		   "N %d %s QP %x R %u %s %u %u f=%x %u %u %u %u %u %u SPSN %x %x %x %x %x RPSN %x S(%u %u %u %u %u %u %u) R(%u %u %u) RQP %x LID %x SL %u MTU %u %u %u %u %u SDE %p,%u SC %p,%u SCQ %u %u PID %d OS %x %x E %x %x %x RNR %d %s %d\n",
616		   iter->n,
617		   qp_idle(qp) ? "I" : "B",
618		   qp->ibqp.qp_num,
619		   atomic_read(&qp->refcount),
620		   qp_type_str[qp->ibqp.qp_type],
621		   qp->state,
622		   wqe ? wqe->wr.opcode : 0,
623		   qp->s_flags,
624		   iowait_sdma_pending(&priv->s_iowait),
625		   iowait_pio_pending(&priv->s_iowait),
626		   !list_empty(&priv->s_iowait.list),
627		   qp->timeout,
628		   wqe ? wqe->ssn : 0,
629		   qp->s_lsn,
630		   qp->s_last_psn,
631		   qp->s_psn, qp->s_next_psn,
632		   qp->s_sending_psn, qp->s_sending_hpsn,
633		   qp->r_psn,
634		   qp->s_last, qp->s_acked, qp->s_cur,
635		   qp->s_tail, qp->s_head, qp->s_size,
636		   qp->s_avail,
637		   /* ack_queue ring pointers, size */
638		   qp->s_tail_ack_queue, qp->r_head_ack_queue,
639		   rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
640		   /* remote QP info  */
641		   qp->remote_qpn,
642		   rdma_ah_get_dlid(&qp->remote_ah_attr),
643		   rdma_ah_get_sl(&qp->remote_ah_attr),
644		   qp->pmtu,
645		   qp->s_retry,
646		   qp->s_retry_cnt,
647		   qp->s_rnr_retry_cnt,
648		   qp->s_rnr_retry,
649		   sde,
650		   sde ? sde->this_idx : 0,
651		   send_context,
652		   send_context ? send_context->sw_index : 0,
653		   ib_cq_head(qp->ibqp.send_cq),
654		   ib_cq_tail(qp->ibqp.send_cq),
655		   qp->pid,
656		   qp->s_state,
657		   qp->s_ack_state,
658		   /* ack queue information */
659		   e ? e->opcode : 0,
660		   e ? e->psn : 0,
661		   e ? e->lpsn : 0,
662		   qp->r_min_rnr_timer,
663		   srq ? "SRQ" : "RQ",
664		   srq ? srq->rq.size : qp->r_rq.size
665		);
666}
667
668void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
669{
670	struct hfi1_qp_priv *priv;
671
672	priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
673	if (!priv)
674		return ERR_PTR(-ENOMEM);
675
676	priv->owner = qp;
677
678	priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
679				   rdi->dparms.node);
680	if (!priv->s_ahg) {
681		kfree(priv);
682		return ERR_PTR(-ENOMEM);
683	}
684	iowait_init(
685		&priv->s_iowait,
686		1,
687		_hfi1_do_send,
688		_hfi1_do_tid_send,
689		iowait_sleep,
690		iowait_wakeup,
691		iowait_sdma_drained,
692		hfi1_init_priority);
693	/* Init to a value to start the running average correctly */
694	priv->s_running_pkt_size = piothreshold / 2;
695	return priv;
696}
697
698void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
699{
700	struct hfi1_qp_priv *priv = qp->priv;
701
702	hfi1_qp_priv_tid_free(rdi, qp);
703	kfree(priv->s_ahg);
704	kfree(priv);
705}
706
707unsigned free_all_qps(struct rvt_dev_info *rdi)
708{
709	struct hfi1_ibdev *verbs_dev = container_of(rdi,
710						    struct hfi1_ibdev,
711						    rdi);
712	struct hfi1_devdata *dd = container_of(verbs_dev,
713					       struct hfi1_devdata,
714					       verbs_dev);
715	int n;
716	unsigned qp_inuse = 0;
717
718	for (n = 0; n < dd->num_pports; n++) {
719		struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
720
721		rcu_read_lock();
722		if (rcu_dereference(ibp->rvp.qp[0]))
723			qp_inuse++;
724		if (rcu_dereference(ibp->rvp.qp[1]))
725			qp_inuse++;
726		rcu_read_unlock();
727	}
728
729	return qp_inuse;
730}
731
732void flush_qp_waiters(struct rvt_qp *qp)
733{
734	lockdep_assert_held(&qp->s_lock);
735	flush_iowait(qp);
736	hfi1_tid_rdma_flush_wait(qp);
737}
738
739void stop_send_queue(struct rvt_qp *qp)
740{
741	struct hfi1_qp_priv *priv = qp->priv;
742
743	iowait_cancel_work(&priv->s_iowait);
744	if (cancel_work_sync(&priv->tid_rdma.trigger_work))
745		rvt_put_qp(qp);
746}
747
748void quiesce_qp(struct rvt_qp *qp)
749{
750	struct hfi1_qp_priv *priv = qp->priv;
751
752	hfi1_del_tid_reap_timer(qp);
753	hfi1_del_tid_retry_timer(qp);
754	iowait_sdma_drain(&priv->s_iowait);
755	qp_pio_drain(qp);
756	flush_tx_list(qp);
757}
758
759void notify_qp_reset(struct rvt_qp *qp)
760{
761	hfi1_qp_kern_exp_rcv_clear_all(qp);
762	qp->r_adefered = 0;
763	clear_ahg(qp);
764
765	/* Clear any OPFN state */
766	if (qp->ibqp.qp_type == IB_QPT_RC)
767		opfn_conn_error(qp);
768}
769
770/*
771 * Switch to alternate path.
772 * The QP s_lock should be held and interrupts disabled.
773 */
774void hfi1_migrate_qp(struct rvt_qp *qp)
775{
776	struct hfi1_qp_priv *priv = qp->priv;
777	struct ib_event ev;
778
779	qp->s_mig_state = IB_MIG_MIGRATED;
780	qp->remote_ah_attr = qp->alt_ah_attr;
781	qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
782	qp->s_pkey_index = qp->s_alt_pkey_index;
783	qp->s_flags |= HFI1_S_AHG_CLEAR;
784	priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
785	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
786	qp_set_16b(qp);
787
788	ev.device = qp->ibqp.device;
789	ev.element.qp = &qp->ibqp;
790	ev.event = IB_EVENT_PATH_MIG;
791	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
792}
793
794int mtu_to_path_mtu(u32 mtu)
795{
796	return mtu_to_enum(mtu, OPA_MTU_8192);
797}
798
799u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
800{
801	u32 mtu;
802	struct hfi1_ibdev *verbs_dev = container_of(rdi,
803						    struct hfi1_ibdev,
804						    rdi);
805	struct hfi1_devdata *dd = container_of(verbs_dev,
806					       struct hfi1_devdata,
807					       verbs_dev);
808	struct hfi1_ibport *ibp;
809	u8 sc, vl;
810
811	ibp = &dd->pport[qp->port_num - 1].ibport_data;
812	sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
813	vl = sc_to_vlt(dd, sc);
814
815	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
816	if (vl < PER_VL_SEND_CONTEXTS)
817		mtu = min_t(u32, mtu, dd->vld[vl].mtu);
818	return mtu;
819}
820
821int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
822		       struct ib_qp_attr *attr)
823{
824	int mtu, pidx = qp->port_num - 1;
825	struct hfi1_ibdev *verbs_dev = container_of(rdi,
826						    struct hfi1_ibdev,
827						    rdi);
828	struct hfi1_devdata *dd = container_of(verbs_dev,
829					       struct hfi1_devdata,
830					       verbs_dev);
831	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
832	if (mtu == -1)
833		return -1; /* values less than 0 are error */
834
835	if (mtu > dd->pport[pidx].ibmtu)
836		return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
837	else
838		return attr->path_mtu;
839}
840
841void notify_error_qp(struct rvt_qp *qp)
842{
843	struct hfi1_qp_priv *priv = qp->priv;
844	seqlock_t *lock = priv->s_iowait.lock;
845
846	if (lock) {
847		write_seqlock(lock);
848		if (!list_empty(&priv->s_iowait.list) &&
849		    !(qp->s_flags & RVT_S_BUSY) &&
850		    !(priv->s_flags & RVT_S_BUSY)) {
851			qp->s_flags &= ~HFI1_S_ANY_WAIT_IO;
852			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
853			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
854			list_del_init(&priv->s_iowait.list);
855			priv->s_iowait.lock = NULL;
856			rvt_put_qp(qp);
857		}
858		write_sequnlock(lock);
859	}
860
861	if (!(qp->s_flags & RVT_S_BUSY) && !(priv->s_flags & RVT_S_BUSY)) {
862		qp->s_hdrwords = 0;
863		if (qp->s_rdma_mr) {
864			rvt_put_mr(qp->s_rdma_mr);
865			qp->s_rdma_mr = NULL;
866		}
867		flush_tx_list(qp);
868	}
869}
870
871/**
872 * hfi1_qp_iter_cb - callback for iterator
873 * @qp: the qp
874 * @v: the sl in low bits of v
875 *
876 * This is called from the iterator callback to work
877 * on an individual qp.
878 */
879static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
880{
881	int lastwqe;
882	struct ib_event ev;
883	struct hfi1_ibport *ibp =
884		to_iport(qp->ibqp.device, qp->port_num);
885	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
886	u8 sl = (u8)v;
887
888	if (qp->port_num != ppd->port ||
889	    (qp->ibqp.qp_type != IB_QPT_UC &&
890	     qp->ibqp.qp_type != IB_QPT_RC) ||
891	    rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
892	    !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
893		return;
894
895	spin_lock_irq(&qp->r_lock);
896	spin_lock(&qp->s_hlock);
897	spin_lock(&qp->s_lock);
898	lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
899	spin_unlock(&qp->s_lock);
900	spin_unlock(&qp->s_hlock);
901	spin_unlock_irq(&qp->r_lock);
902	if (lastwqe) {
903		ev.device = qp->ibqp.device;
904		ev.element.qp = &qp->ibqp;
905		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
906		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
907	}
908}
909
910/**
911 * hfi1_error_port_qps - put a port's RC/UC qps into error state
912 * @ibp: the ibport.
913 * @sl: the service level.
914 *
915 * This function places all RC/UC qps with a given service level into error
916 * state. It is generally called to force upper lay apps to abandon stale qps
917 * after an sl->sc mapping change.
918 */
919void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
920{
921	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
922	struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
923
924	rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);
925}
v5.4
 
  1/*
  2 * Copyright(c) 2015 - 2019 Intel Corporation.
  3 *
  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.
  6 *
  7 * GPL LICENSE SUMMARY
  8 *
  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.
 12 *
 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.
 17 *
 18 * BSD LICENSE
 19 *
 20 * Redistribution and use in source and binary forms, with or without
 21 * modification, are permitted provided that the following conditions
 22 * are met:
 23 *
 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
 29 *    distribution.
 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.
 33 *
 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.
 45 *
 46 */
 47
 48#include <linux/err.h>
 49#include <linux/vmalloc.h>
 50#include <linux/hash.h>
 51#include <linux/module.h>
 52#include <linux/seq_file.h>
 53#include <rdma/rdma_vt.h>
 54#include <rdma/rdmavt_qp.h>
 55#include <rdma/ib_verbs.h>
 56
 57#include "hfi.h"
 58#include "qp.h"
 59#include "trace.h"
 60#include "verbs_txreq.h"
 61
 62unsigned int hfi1_qp_table_size = 256;
 63module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO);
 64MODULE_PARM_DESC(qp_table_size, "QP table size");
 65
 66static void flush_tx_list(struct rvt_qp *qp);
 67static int iowait_sleep(
 68	struct sdma_engine *sde,
 69	struct iowait_work *wait,
 70	struct sdma_txreq *stx,
 71	unsigned int seq,
 72	bool pkts_sent);
 73static void iowait_wakeup(struct iowait *wait, int reason);
 74static void iowait_sdma_drained(struct iowait *wait);
 75static void qp_pio_drain(struct rvt_qp *qp);
 76
 77const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
 78[IB_WR_RDMA_WRITE] = {
 79	.length = sizeof(struct ib_rdma_wr),
 80	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 81},
 82
 83[IB_WR_RDMA_READ] = {
 84	.length = sizeof(struct ib_rdma_wr),
 85	.qpt_support = BIT(IB_QPT_RC),
 86	.flags = RVT_OPERATION_ATOMIC,
 87},
 88
 89[IB_WR_ATOMIC_CMP_AND_SWP] = {
 90	.length = sizeof(struct ib_atomic_wr),
 91	.qpt_support = BIT(IB_QPT_RC),
 92	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
 93},
 94
 95[IB_WR_ATOMIC_FETCH_AND_ADD] = {
 96	.length = sizeof(struct ib_atomic_wr),
 97	.qpt_support = BIT(IB_QPT_RC),
 98	.flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
 99},
100
101[IB_WR_RDMA_WRITE_WITH_IMM] = {
102	.length = sizeof(struct ib_rdma_wr),
103	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
104},
105
106[IB_WR_SEND] = {
107	.length = sizeof(struct ib_send_wr),
108	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
109		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
110},
111
112[IB_WR_SEND_WITH_IMM] = {
113	.length = sizeof(struct ib_send_wr),
114	.qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
115		       BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
116},
117
118[IB_WR_REG_MR] = {
119	.length = sizeof(struct ib_reg_wr),
120	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
121	.flags = RVT_OPERATION_LOCAL,
122},
123
124[IB_WR_LOCAL_INV] = {
125	.length = sizeof(struct ib_send_wr),
126	.qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
127	.flags = RVT_OPERATION_LOCAL,
128},
129
130[IB_WR_SEND_WITH_INV] = {
131	.length = sizeof(struct ib_send_wr),
132	.qpt_support = BIT(IB_QPT_RC),
133},
134
135[IB_WR_OPFN] = {
136	.length = sizeof(struct ib_atomic_wr),
137	.qpt_support = BIT(IB_QPT_RC),
138	.flags = RVT_OPERATION_USE_RESERVE,
139},
140
141[IB_WR_TID_RDMA_WRITE] = {
142	.length = sizeof(struct ib_rdma_wr),
143	.qpt_support = BIT(IB_QPT_RC),
144	.flags = RVT_OPERATION_IGN_RNR_CNT,
145},
146
147};
148
149static void flush_list_head(struct list_head *l)
150{
151	while (!list_empty(l)) {
152		struct sdma_txreq *tx;
153
154		tx = list_first_entry(
155			l,
156			struct sdma_txreq,
157			list);
158		list_del_init(&tx->list);
159		hfi1_put_txreq(
160			container_of(tx, struct verbs_txreq, txreq));
161	}
162}
163
164static void flush_tx_list(struct rvt_qp *qp)
165{
166	struct hfi1_qp_priv *priv = qp->priv;
167
168	flush_list_head(&iowait_get_ib_work(&priv->s_iowait)->tx_head);
169	flush_list_head(&iowait_get_tid_work(&priv->s_iowait)->tx_head);
170}
171
172static void flush_iowait(struct rvt_qp *qp)
173{
174	struct hfi1_qp_priv *priv = qp->priv;
175	unsigned long flags;
176	seqlock_t *lock = priv->s_iowait.lock;
177
178	if (!lock)
179		return;
180	write_seqlock_irqsave(lock, flags);
181	if (!list_empty(&priv->s_iowait.list)) {
182		list_del_init(&priv->s_iowait.list);
183		priv->s_iowait.lock = NULL;
184		rvt_put_qp(qp);
185	}
186	write_sequnlock_irqrestore(lock, flags);
187}
188
189static inline int opa_mtu_enum_to_int(int mtu)
190{
191	switch (mtu) {
192	case OPA_MTU_8192:  return 8192;
193	case OPA_MTU_10240: return 10240;
194	default:            return -1;
195	}
196}
197
198/**
199 * This function is what we would push to the core layer if we wanted to be a
200 * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
201 * to blindly pass the MTU enum value from the PathRecord to us.
202 */
203static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
204{
205	int val;
206
207	/* Constraining 10KB packets to 8KB packets */
208	if (mtu == (enum ib_mtu)OPA_MTU_10240)
209		mtu = OPA_MTU_8192;
210	val = opa_mtu_enum_to_int((int)mtu);
211	if (val > 0)
212		return val;
213	return ib_mtu_enum_to_int(mtu);
214}
215
216int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
217			 int attr_mask, struct ib_udata *udata)
218{
219	struct ib_qp *ibqp = &qp->ibqp;
220	struct hfi1_ibdev *dev = to_idev(ibqp->device);
221	struct hfi1_devdata *dd = dd_from_dev(dev);
222	u8 sc;
223
224	if (attr_mask & IB_QP_AV) {
225		sc = ah_to_sc(ibqp->device, &attr->ah_attr);
226		if (sc == 0xf)
227			return -EINVAL;
228
229		if (!qp_to_sdma_engine(qp, sc) &&
230		    dd->flags & HFI1_HAS_SEND_DMA)
231			return -EINVAL;
232
233		if (!qp_to_send_context(qp, sc))
234			return -EINVAL;
235	}
236
237	if (attr_mask & IB_QP_ALT_PATH) {
238		sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
239		if (sc == 0xf)
240			return -EINVAL;
241
242		if (!qp_to_sdma_engine(qp, sc) &&
243		    dd->flags & HFI1_HAS_SEND_DMA)
244			return -EINVAL;
245
246		if (!qp_to_send_context(qp, sc))
247			return -EINVAL;
248	}
249
250	return 0;
251}
252
253/*
254 * qp_set_16b - Set the hdr_type based on whether the slid or the
255 * dlid in the connection is extended. Only applicable for RC and UC
256 * QPs. UD QPs determine this on the fly from the ah in the wqe
257 */
258static inline void qp_set_16b(struct rvt_qp *qp)
259{
260	struct hfi1_pportdata *ppd;
261	struct hfi1_ibport *ibp;
262	struct hfi1_qp_priv *priv = qp->priv;
263
264	/* Update ah_attr to account for extended LIDs */
265	hfi1_update_ah_attr(qp->ibqp.device, &qp->remote_ah_attr);
266
267	/* Create 32 bit LIDs */
268	hfi1_make_opa_lid(&qp->remote_ah_attr);
269
270	if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH))
271		return;
272
273	ibp = to_iport(qp->ibqp.device, qp->port_num);
274	ppd = ppd_from_ibp(ibp);
275	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, &qp->remote_ah_attr);
276}
277
278void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
279		    int attr_mask, struct ib_udata *udata)
280{
281	struct ib_qp *ibqp = &qp->ibqp;
282	struct hfi1_qp_priv *priv = qp->priv;
283
284	if (attr_mask & IB_QP_AV) {
285		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
286		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
287		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
288		qp_set_16b(qp);
289	}
290
291	if (attr_mask & IB_QP_PATH_MIG_STATE &&
292	    attr->path_mig_state == IB_MIG_MIGRATED &&
293	    qp->s_mig_state == IB_MIG_ARMED) {
294		qp->s_flags |= HFI1_S_AHG_CLEAR;
295		priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
296		priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
297		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
298		qp_set_16b(qp);
299	}
300
301	opfn_qp_init(qp, attr, attr_mask);
302}
303
304/**
305 * hfi1_setup_wqe - set up the wqe
306 * @qp - The qp
307 * @wqe - The built wqe
308 * @call_send - Determine if the send should be posted or scheduled.
309 *
310 * Perform setup of the wqe.  This is called
311 * prior to inserting the wqe into the ring but after
312 * the wqe has been setup by RDMAVT. This function
313 * allows the driver the opportunity to perform
314 * validation and additional setup of the wqe.
315 *
316 * Returns 0 on success, -EINVAL on failure
317 *
318 */
319int hfi1_setup_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe, bool *call_send)
320{
321	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
322	struct rvt_ah *ah;
323	struct hfi1_pportdata *ppd;
324	struct hfi1_devdata *dd;
325
326	switch (qp->ibqp.qp_type) {
327	case IB_QPT_RC:
328		hfi1_setup_tid_rdma_wqe(qp, wqe);
329		/* fall through */
330	case IB_QPT_UC:
331		if (wqe->length > 0x80000000U)
332			return -EINVAL;
333		if (wqe->length > qp->pmtu)
334			*call_send = false;
335		break;
336	case IB_QPT_SMI:
337		/*
338		 * SM packets should exclusively use VL15 and their SL is
339		 * ignored (IBTA v1.3, Section 3.5.8.2). Therefore, when ah
340		 * is created, SL is 0 in most cases and as a result some
341		 * fields (vl and pmtu) in ah may not be set correctly,
342		 * depending on the SL2SC and SC2VL tables at the time.
343		 */
344		ppd = ppd_from_ibp(ibp);
345		dd = dd_from_ppd(ppd);
346		if (wqe->length > dd->vld[15].mtu)
347			return -EINVAL;
348		break;
349	case IB_QPT_GSI:
350	case IB_QPT_UD:
351		ah = rvt_get_swqe_ah(wqe);
352		if (wqe->length > (1 << ah->log_pmtu))
353			return -EINVAL;
354		if (ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)] == 0xf)
355			return -EINVAL;
 
356	default:
357		break;
358	}
359
360	/*
361	 * System latency between send and schedule is large enough that
362	 * forcing call_send to true for piothreshold packets is necessary.
363	 */
364	if (wqe->length <= piothreshold)
365		*call_send = true;
366	return 0;
367}
368
369/**
370 * _hfi1_schedule_send - schedule progress
371 * @qp: the QP
372 *
373 * This schedules qp progress w/o regard to the s_flags.
374 *
375 * It is only used in the post send, which doesn't hold
376 * the s_lock.
377 */
378bool _hfi1_schedule_send(struct rvt_qp *qp)
379{
380	struct hfi1_qp_priv *priv = qp->priv;
381	struct hfi1_ibport *ibp =
382		to_iport(qp->ibqp.device, qp->port_num);
383	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
384	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 
 
 
385
386	return iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
387			       priv->s_sde ?
388			       priv->s_sde->cpu :
389			       cpumask_first(cpumask_of_node(dd->node)));
390}
391
392static void qp_pio_drain(struct rvt_qp *qp)
393{
394	struct hfi1_qp_priv *priv = qp->priv;
395
396	if (!priv->s_sendcontext)
397		return;
398	while (iowait_pio_pending(&priv->s_iowait)) {
399		write_seqlock_irq(&priv->s_sendcontext->waitlock);
400		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
401		write_sequnlock_irq(&priv->s_sendcontext->waitlock);
402		iowait_pio_drain(&priv->s_iowait);
403		write_seqlock_irq(&priv->s_sendcontext->waitlock);
404		hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
405		write_sequnlock_irq(&priv->s_sendcontext->waitlock);
406	}
407}
408
409/**
410 * hfi1_schedule_send - schedule progress
411 * @qp: the QP
412 *
413 * This schedules qp progress and caller should hold
414 * the s_lock.
415 * @return true if the first leg is scheduled;
416 * false if the first leg is not scheduled.
417 */
418bool hfi1_schedule_send(struct rvt_qp *qp)
419{
420	lockdep_assert_held(&qp->s_lock);
421	if (hfi1_send_ok(qp)) {
422		_hfi1_schedule_send(qp);
423		return true;
424	}
425	if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
426		iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
427				IOWAIT_PENDING_IB);
428	return false;
429}
430
431static void hfi1_qp_schedule(struct rvt_qp *qp)
432{
433	struct hfi1_qp_priv *priv = qp->priv;
434	bool ret;
435
436	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_IB)) {
437		ret = hfi1_schedule_send(qp);
438		if (ret)
439			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
440	}
441	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_TID)) {
442		ret = hfi1_schedule_tid_send(qp);
443		if (ret)
444			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
445	}
446}
447
448void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
449{
450	unsigned long flags;
451
452	spin_lock_irqsave(&qp->s_lock, flags);
453	if (qp->s_flags & flag) {
454		qp->s_flags &= ~flag;
455		trace_hfi1_qpwakeup(qp, flag);
456		hfi1_qp_schedule(qp);
457	}
458	spin_unlock_irqrestore(&qp->s_lock, flags);
459	/* Notify hfi1_destroy_qp() if it is waiting. */
460	rvt_put_qp(qp);
461}
462
463void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
464{
465	struct hfi1_qp_priv *priv = qp->priv;
466
467	if (iowait_set_work_flag(wait) == IOWAIT_IB_SE) {
468		qp->s_flags &= ~RVT_S_BUSY;
469		/*
470		 * If we are sending a first-leg packet from the second leg,
471		 * we need to clear the busy flag from priv->s_flags to
472		 * avoid a race condition when the qp wakes up before
473		 * the call to hfi1_verbs_send() returns to the second
474		 * leg. In that case, the second leg will terminate without
475		 * being re-scheduled, resulting in failure to send TID RDMA
476		 * WRITE DATA and TID RDMA ACK packets.
477		 */
478		if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
479			priv->s_flags &= ~(HFI1_S_TID_BUSY_SET |
480					   RVT_S_BUSY);
481			iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
482		}
483	} else {
484		priv->s_flags &= ~RVT_S_BUSY;
485	}
486}
487
488static int iowait_sleep(
489	struct sdma_engine *sde,
490	struct iowait_work *wait,
491	struct sdma_txreq *stx,
492	uint seq,
493	bool pkts_sent)
494{
495	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
496	struct rvt_qp *qp;
497	struct hfi1_qp_priv *priv;
498	unsigned long flags;
499	int ret = 0;
500
501	qp = tx->qp;
502	priv = qp->priv;
503
504	spin_lock_irqsave(&qp->s_lock, flags);
505	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
506		/*
507		 * If we couldn't queue the DMA request, save the info
508		 * and try again later rather than destroying the
509		 * buffer and undoing the side effects of the copy.
510		 */
511		/* Make a common routine? */
512		list_add_tail(&stx->list, &wait->tx_head);
513		write_seqlock(&sde->waitlock);
514		if (sdma_progress(sde, seq, stx))
515			goto eagain;
516		if (list_empty(&priv->s_iowait.list)) {
517			struct hfi1_ibport *ibp =
518				to_iport(qp->ibqp.device, qp->port_num);
519
520			ibp->rvp.n_dmawait++;
521			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
522			iowait_get_priority(&priv->s_iowait);
523			iowait_queue(pkts_sent, &priv->s_iowait,
524				     &sde->dmawait);
525			priv->s_iowait.lock = &sde->waitlock;
526			trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
527			rvt_get_qp(qp);
528		}
529		write_sequnlock(&sde->waitlock);
530		hfi1_qp_unbusy(qp, wait);
531		spin_unlock_irqrestore(&qp->s_lock, flags);
532		ret = -EBUSY;
533	} else {
534		spin_unlock_irqrestore(&qp->s_lock, flags);
535		hfi1_put_txreq(tx);
536	}
537	return ret;
538eagain:
539	write_sequnlock(&sde->waitlock);
540	spin_unlock_irqrestore(&qp->s_lock, flags);
541	list_del_init(&stx->list);
542	return -EAGAIN;
543}
544
545static void iowait_wakeup(struct iowait *wait, int reason)
546{
547	struct rvt_qp *qp = iowait_to_qp(wait);
548
549	WARN_ON(reason != SDMA_AVAIL_REASON);
550	hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
551}
552
553static void iowait_sdma_drained(struct iowait *wait)
554{
555	struct rvt_qp *qp = iowait_to_qp(wait);
556	unsigned long flags;
557
558	/*
559	 * This happens when the send engine notes
560	 * a QP in the error state and cannot
561	 * do the flush work until that QP's
562	 * sdma work has finished.
563	 */
564	spin_lock_irqsave(&qp->s_lock, flags);
565	if (qp->s_flags & RVT_S_WAIT_DMA) {
566		qp->s_flags &= ~RVT_S_WAIT_DMA;
567		hfi1_schedule_send(qp);
568	}
569	spin_unlock_irqrestore(&qp->s_lock, flags);
570}
571
572static void hfi1_init_priority(struct iowait *w)
573{
574	struct rvt_qp *qp = iowait_to_qp(w);
575	struct hfi1_qp_priv *priv = qp->priv;
576
577	if (qp->s_flags & RVT_S_ACK_PENDING)
578		w->priority++;
579	if (priv->s_flags & RVT_S_ACK_PENDING)
580		w->priority++;
581}
582
583/**
584 * qp_to_sdma_engine - map a qp to a send engine
585 * @qp: the QP
586 * @sc5: the 5 bit sc
587 *
588 * Return:
589 * A send engine for the qp or NULL for SMI type qp.
590 */
591struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
592{
593	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
594	struct sdma_engine *sde;
595
596	if (!(dd->flags & HFI1_HAS_SEND_DMA))
597		return NULL;
598	switch (qp->ibqp.qp_type) {
599	case IB_QPT_SMI:
600		return NULL;
601	default:
602		break;
603	}
604	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
605	return sde;
606}
607
608/*
609 * qp_to_send_context - map a qp to a send context
610 * @qp: the QP
611 * @sc5: the 5 bit sc
612 *
613 * Return:
614 * A send context for the qp
615 */
616struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
617{
618	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
619
620	switch (qp->ibqp.qp_type) {
621	case IB_QPT_SMI:
622		/* SMA packets to VL15 */
623		return dd->vld[15].sc;
624	default:
625		break;
626	}
627
628	return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
629					  sc5);
630}
631
632static const char * const qp_type_str[] = {
633	"SMI", "GSI", "RC", "UC", "UD",
634};
635
636static int qp_idle(struct rvt_qp *qp)
637{
638	return
639		qp->s_last == qp->s_acked &&
640		qp->s_acked == qp->s_cur &&
641		qp->s_cur == qp->s_tail &&
642		qp->s_tail == qp->s_head;
643}
644
645/**
646 * qp_iter_print - print the qp information to seq_file
647 * @s: the seq_file to emit the qp information on
648 * @iter: the iterator for the qp hash list
649 */
650void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
651{
652	struct rvt_swqe *wqe;
653	struct rvt_qp *qp = iter->qp;
654	struct hfi1_qp_priv *priv = qp->priv;
655	struct sdma_engine *sde;
656	struct send_context *send_context;
657	struct rvt_ack_entry *e = NULL;
658	struct rvt_srq *srq = qp->ibqp.srq ?
659		ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
660
661	sde = qp_to_sdma_engine(qp, priv->s_sc);
662	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
663	send_context = qp_to_send_context(qp, priv->s_sc);
664	if (qp->s_ack_queue)
665		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
666	seq_printf(s,
667		   "N %d %s QP %x R %u %s %u %u f=%x %u %u %u %u %u %u SPSN %x %x %x %x %x RPSN %x S(%u %u %u %u %u %u %u) R(%u %u %u) RQP %x LID %x SL %u MTU %u %u %u %u %u SDE %p,%u SC %p,%u SCQ %u %u PID %d OS %x %x E %x %x %x RNR %d %s %d\n",
668		   iter->n,
669		   qp_idle(qp) ? "I" : "B",
670		   qp->ibqp.qp_num,
671		   atomic_read(&qp->refcount),
672		   qp_type_str[qp->ibqp.qp_type],
673		   qp->state,
674		   wqe ? wqe->wr.opcode : 0,
675		   qp->s_flags,
676		   iowait_sdma_pending(&priv->s_iowait),
677		   iowait_pio_pending(&priv->s_iowait),
678		   !list_empty(&priv->s_iowait.list),
679		   qp->timeout,
680		   wqe ? wqe->ssn : 0,
681		   qp->s_lsn,
682		   qp->s_last_psn,
683		   qp->s_psn, qp->s_next_psn,
684		   qp->s_sending_psn, qp->s_sending_hpsn,
685		   qp->r_psn,
686		   qp->s_last, qp->s_acked, qp->s_cur,
687		   qp->s_tail, qp->s_head, qp->s_size,
688		   qp->s_avail,
689		   /* ack_queue ring pointers, size */
690		   qp->s_tail_ack_queue, qp->r_head_ack_queue,
691		   rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
692		   /* remote QP info  */
693		   qp->remote_qpn,
694		   rdma_ah_get_dlid(&qp->remote_ah_attr),
695		   rdma_ah_get_sl(&qp->remote_ah_attr),
696		   qp->pmtu,
697		   qp->s_retry,
698		   qp->s_retry_cnt,
699		   qp->s_rnr_retry_cnt,
700		   qp->s_rnr_retry,
701		   sde,
702		   sde ? sde->this_idx : 0,
703		   send_context,
704		   send_context ? send_context->sw_index : 0,
705		   ib_cq_head(qp->ibqp.send_cq),
706		   ib_cq_tail(qp->ibqp.send_cq),
707		   qp->pid,
708		   qp->s_state,
709		   qp->s_ack_state,
710		   /* ack queue information */
711		   e ? e->opcode : 0,
712		   e ? e->psn : 0,
713		   e ? e->lpsn : 0,
714		   qp->r_min_rnr_timer,
715		   srq ? "SRQ" : "RQ",
716		   srq ? srq->rq.size : qp->r_rq.size
717		);
718}
719
720void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
721{
722	struct hfi1_qp_priv *priv;
723
724	priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
725	if (!priv)
726		return ERR_PTR(-ENOMEM);
727
728	priv->owner = qp;
729
730	priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
731				   rdi->dparms.node);
732	if (!priv->s_ahg) {
733		kfree(priv);
734		return ERR_PTR(-ENOMEM);
735	}
736	iowait_init(
737		&priv->s_iowait,
738		1,
739		_hfi1_do_send,
740		_hfi1_do_tid_send,
741		iowait_sleep,
742		iowait_wakeup,
743		iowait_sdma_drained,
744		hfi1_init_priority);
745	/* Init to a value to start the running average correctly */
746	priv->s_running_pkt_size = piothreshold / 2;
747	return priv;
748}
749
750void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
751{
752	struct hfi1_qp_priv *priv = qp->priv;
753
754	hfi1_qp_priv_tid_free(rdi, qp);
755	kfree(priv->s_ahg);
756	kfree(priv);
757}
758
759unsigned free_all_qps(struct rvt_dev_info *rdi)
760{
761	struct hfi1_ibdev *verbs_dev = container_of(rdi,
762						    struct hfi1_ibdev,
763						    rdi);
764	struct hfi1_devdata *dd = container_of(verbs_dev,
765					       struct hfi1_devdata,
766					       verbs_dev);
767	int n;
768	unsigned qp_inuse = 0;
769
770	for (n = 0; n < dd->num_pports; n++) {
771		struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
772
773		rcu_read_lock();
774		if (rcu_dereference(ibp->rvp.qp[0]))
775			qp_inuse++;
776		if (rcu_dereference(ibp->rvp.qp[1]))
777			qp_inuse++;
778		rcu_read_unlock();
779	}
780
781	return qp_inuse;
782}
783
784void flush_qp_waiters(struct rvt_qp *qp)
785{
786	lockdep_assert_held(&qp->s_lock);
787	flush_iowait(qp);
788	hfi1_tid_rdma_flush_wait(qp);
789}
790
791void stop_send_queue(struct rvt_qp *qp)
792{
793	struct hfi1_qp_priv *priv = qp->priv;
794
795	iowait_cancel_work(&priv->s_iowait);
796	if (cancel_work_sync(&priv->tid_rdma.trigger_work))
797		rvt_put_qp(qp);
798}
799
800void quiesce_qp(struct rvt_qp *qp)
801{
802	struct hfi1_qp_priv *priv = qp->priv;
803
804	hfi1_del_tid_reap_timer(qp);
805	hfi1_del_tid_retry_timer(qp);
806	iowait_sdma_drain(&priv->s_iowait);
807	qp_pio_drain(qp);
808	flush_tx_list(qp);
809}
810
811void notify_qp_reset(struct rvt_qp *qp)
812{
813	hfi1_qp_kern_exp_rcv_clear_all(qp);
814	qp->r_adefered = 0;
815	clear_ahg(qp);
816
817	/* Clear any OPFN state */
818	if (qp->ibqp.qp_type == IB_QPT_RC)
819		opfn_conn_error(qp);
820}
821
822/*
823 * Switch to alternate path.
824 * The QP s_lock should be held and interrupts disabled.
825 */
826void hfi1_migrate_qp(struct rvt_qp *qp)
827{
828	struct hfi1_qp_priv *priv = qp->priv;
829	struct ib_event ev;
830
831	qp->s_mig_state = IB_MIG_MIGRATED;
832	qp->remote_ah_attr = qp->alt_ah_attr;
833	qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
834	qp->s_pkey_index = qp->s_alt_pkey_index;
835	qp->s_flags |= HFI1_S_AHG_CLEAR;
836	priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
837	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
838	qp_set_16b(qp);
839
840	ev.device = qp->ibqp.device;
841	ev.element.qp = &qp->ibqp;
842	ev.event = IB_EVENT_PATH_MIG;
843	qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
844}
845
846int mtu_to_path_mtu(u32 mtu)
847{
848	return mtu_to_enum(mtu, OPA_MTU_8192);
849}
850
851u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
852{
853	u32 mtu;
854	struct hfi1_ibdev *verbs_dev = container_of(rdi,
855						    struct hfi1_ibdev,
856						    rdi);
857	struct hfi1_devdata *dd = container_of(verbs_dev,
858					       struct hfi1_devdata,
859					       verbs_dev);
860	struct hfi1_ibport *ibp;
861	u8 sc, vl;
862
863	ibp = &dd->pport[qp->port_num - 1].ibport_data;
864	sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
865	vl = sc_to_vlt(dd, sc);
866
867	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
868	if (vl < PER_VL_SEND_CONTEXTS)
869		mtu = min_t(u32, mtu, dd->vld[vl].mtu);
870	return mtu;
871}
872
873int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
874		       struct ib_qp_attr *attr)
875{
876	int mtu, pidx = qp->port_num - 1;
877	struct hfi1_ibdev *verbs_dev = container_of(rdi,
878						    struct hfi1_ibdev,
879						    rdi);
880	struct hfi1_devdata *dd = container_of(verbs_dev,
881					       struct hfi1_devdata,
882					       verbs_dev);
883	mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
884	if (mtu == -1)
885		return -1; /* values less than 0 are error */
886
887	if (mtu > dd->pport[pidx].ibmtu)
888		return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
889	else
890		return attr->path_mtu;
891}
892
893void notify_error_qp(struct rvt_qp *qp)
894{
895	struct hfi1_qp_priv *priv = qp->priv;
896	seqlock_t *lock = priv->s_iowait.lock;
897
898	if (lock) {
899		write_seqlock(lock);
900		if (!list_empty(&priv->s_iowait.list) &&
901		    !(qp->s_flags & RVT_S_BUSY) &&
902		    !(priv->s_flags & RVT_S_BUSY)) {
903			qp->s_flags &= ~HFI1_S_ANY_WAIT_IO;
904			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
905			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
906			list_del_init(&priv->s_iowait.list);
907			priv->s_iowait.lock = NULL;
908			rvt_put_qp(qp);
909		}
910		write_sequnlock(lock);
911	}
912
913	if (!(qp->s_flags & RVT_S_BUSY) && !(priv->s_flags & RVT_S_BUSY)) {
914		qp->s_hdrwords = 0;
915		if (qp->s_rdma_mr) {
916			rvt_put_mr(qp->s_rdma_mr);
917			qp->s_rdma_mr = NULL;
918		}
919		flush_tx_list(qp);
920	}
921}
922
923/**
924 * hfi1_qp_iter_cb - callback for iterator
925 * @qp - the qp
926 * @v - the sl in low bits of v
927 *
928 * This is called from the iterator callback to work
929 * on an individual qp.
930 */
931static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
932{
933	int lastwqe;
934	struct ib_event ev;
935	struct hfi1_ibport *ibp =
936		to_iport(qp->ibqp.device, qp->port_num);
937	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
938	u8 sl = (u8)v;
939
940	if (qp->port_num != ppd->port ||
941	    (qp->ibqp.qp_type != IB_QPT_UC &&
942	     qp->ibqp.qp_type != IB_QPT_RC) ||
943	    rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
944	    !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
945		return;
946
947	spin_lock_irq(&qp->r_lock);
948	spin_lock(&qp->s_hlock);
949	spin_lock(&qp->s_lock);
950	lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
951	spin_unlock(&qp->s_lock);
952	spin_unlock(&qp->s_hlock);
953	spin_unlock_irq(&qp->r_lock);
954	if (lastwqe) {
955		ev.device = qp->ibqp.device;
956		ev.element.qp = &qp->ibqp;
957		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
958		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
959	}
960}
961
962/**
963 * hfi1_error_port_qps - put a port's RC/UC qps into error state
964 * @ibp: the ibport.
965 * @sl: the service level.
966 *
967 * This function places all RC/UC qps with a given service level into error
968 * state. It is generally called to force upper lay apps to abandon stale qps
969 * after an sl->sc mapping change.
970 */
971void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
972{
973	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
974	struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
975
976	rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);
977}