1 // SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause)
3 * Copyright(c) 2018 Intel Corporation.
16 * DOC: TID RDMA READ protocol
18 * This is an end-to-end protocol at the hfi1 level between two nodes that
19 * improves performance by avoiding data copy on the requester side. It
20 * converts a qualified RDMA READ request into a TID RDMA READ request on
21 * the requester side and thereafter handles the request and response
22 * differently. To be qualified, the RDMA READ request should meet the
24 * -- The total data length should be greater than 256K;
25 * -- The total data length should be a multiple of 4K page size;
26 * -- Each local scatter-gather entry should be 4K page aligned;
27 * -- Each local scatter-gather entry should be a multiple of 4K page size;
30 #define RCV_TID_FLOW_TABLE_CTRL_FLOW_VALID_SMASK BIT_ULL(32)
31 #define RCV_TID_FLOW_TABLE_CTRL_HDR_SUPP_EN_SMASK BIT_ULL(33)
32 #define RCV_TID_FLOW_TABLE_CTRL_KEEP_AFTER_SEQ_ERR_SMASK BIT_ULL(34)
33 #define RCV_TID_FLOW_TABLE_CTRL_KEEP_ON_GEN_ERR_SMASK BIT_ULL(35)
34 #define RCV_TID_FLOW_TABLE_STATUS_SEQ_MISMATCH_SMASK BIT_ULL(37)
35 #define RCV_TID_FLOW_TABLE_STATUS_GEN_MISMATCH_SMASK BIT_ULL(38)
37 /* Maximum number of packets within a flow generation. */
38 #define MAX_TID_FLOW_PSN BIT(HFI1_KDETH_BTH_SEQ_SHIFT)
40 #define GENERATION_MASK 0xFFFFF
42 static u32 mask_generation(u32 a)
44 return a & GENERATION_MASK;
47 /* Reserved generation value to set to unused flows for kernel contexts */
48 #define KERN_GENERATION_RESERVED mask_generation(U32_MAX)
51 * J_KEY for kernel contexts when TID RDMA is used.
52 * See generate_jkey() in hfi.h for more information.
54 #define TID_RDMA_JKEY 32
55 #define HFI1_KERNEL_MIN_JKEY HFI1_ADMIN_JKEY_RANGE
56 #define HFI1_KERNEL_MAX_JKEY (2 * HFI1_ADMIN_JKEY_RANGE - 1)
58 /* Maximum number of segments in flight per QP request. */
59 #define TID_RDMA_MAX_READ_SEGS_PER_REQ 6
60 #define TID_RDMA_MAX_WRITE_SEGS_PER_REQ 4
61 #define MAX_REQ max_t(u16, TID_RDMA_MAX_READ_SEGS_PER_REQ, \
62 TID_RDMA_MAX_WRITE_SEGS_PER_REQ)
63 #define MAX_FLOWS roundup_pow_of_two(MAX_REQ + 1)
65 #define MAX_EXPECTED_PAGES (MAX_EXPECTED_BUFFER / PAGE_SIZE)
67 #define TID_RDMA_DESTQP_FLOW_SHIFT 11
68 #define TID_RDMA_DESTQP_FLOW_MASK 0x1f
70 #define TID_OPFN_QP_CTXT_MASK 0xff
71 #define TID_OPFN_QP_CTXT_SHIFT 56
72 #define TID_OPFN_QP_KDETH_MASK 0xff
73 #define TID_OPFN_QP_KDETH_SHIFT 48
74 #define TID_OPFN_MAX_LEN_MASK 0x7ff
75 #define TID_OPFN_MAX_LEN_SHIFT 37
76 #define TID_OPFN_TIMEOUT_MASK 0x1f
77 #define TID_OPFN_TIMEOUT_SHIFT 32
78 #define TID_OPFN_RESERVED_MASK 0x3f
79 #define TID_OPFN_RESERVED_SHIFT 26
80 #define TID_OPFN_URG_MASK 0x1
81 #define TID_OPFN_URG_SHIFT 25
82 #define TID_OPFN_VER_MASK 0x7
83 #define TID_OPFN_VER_SHIFT 22
84 #define TID_OPFN_JKEY_MASK 0x3f
85 #define TID_OPFN_JKEY_SHIFT 16
86 #define TID_OPFN_MAX_READ_MASK 0x3f
87 #define TID_OPFN_MAX_READ_SHIFT 10
88 #define TID_OPFN_MAX_WRITE_MASK 0x3f
89 #define TID_OPFN_MAX_WRITE_SHIFT 4
95 * NNNNNNNNKKKKKKKK MMMMMMMMMMMTTTTT DDDDDDUVVVJJJJJJ RRRRRRWWWWWWCCCC
96 * 3210987654321098 7654321098765432 1098765432109876 5432109876543210
97 * N - the context Number
110 static u32 tid_rdma_flow_wt;
112 static void tid_rdma_trigger_resume(struct work_struct *work);
113 static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req);
114 static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req,
116 static void hfi1_init_trdma_req(struct rvt_qp *qp,
117 struct tid_rdma_request *req);
118 static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx);
119 static void hfi1_tid_timeout(struct timer_list *t);
120 static void hfi1_add_tid_reap_timer(struct rvt_qp *qp);
121 static void hfi1_mod_tid_reap_timer(struct rvt_qp *qp);
122 static void hfi1_mod_tid_retry_timer(struct rvt_qp *qp);
123 static int hfi1_stop_tid_retry_timer(struct rvt_qp *qp);
124 static void hfi1_tid_retry_timeout(struct timer_list *t);
125 static int make_tid_rdma_ack(struct rvt_qp *qp,
126 struct ib_other_headers *ohdr,
127 struct hfi1_pkt_state *ps);
128 static void hfi1_do_tid_send(struct rvt_qp *qp);
129 static u32 read_r_next_psn(struct hfi1_devdata *dd, u8 ctxt, u8 fidx);
130 static void tid_rdma_rcv_err(struct hfi1_packet *packet,
131 struct ib_other_headers *ohdr,
132 struct rvt_qp *qp, u32 psn, int diff, bool fecn);
133 static void update_r_next_psn_fecn(struct hfi1_packet *packet,
134 struct hfi1_qp_priv *priv,
135 struct hfi1_ctxtdata *rcd,
136 struct tid_rdma_flow *flow,
139 static u64 tid_rdma_opfn_encode(struct tid_rdma_params *p)
142 (((u64)p->qp & TID_OPFN_QP_CTXT_MASK) <<
143 TID_OPFN_QP_CTXT_SHIFT) |
144 ((((u64)p->qp >> 16) & TID_OPFN_QP_KDETH_MASK) <<
145 TID_OPFN_QP_KDETH_SHIFT) |
146 (((u64)((p->max_len >> PAGE_SHIFT) - 1) &
147 TID_OPFN_MAX_LEN_MASK) << TID_OPFN_MAX_LEN_SHIFT) |
148 (((u64)p->timeout & TID_OPFN_TIMEOUT_MASK) <<
149 TID_OPFN_TIMEOUT_SHIFT) |
150 (((u64)p->urg & TID_OPFN_URG_MASK) << TID_OPFN_URG_SHIFT) |
151 (((u64)p->jkey & TID_OPFN_JKEY_MASK) << TID_OPFN_JKEY_SHIFT) |
152 (((u64)p->max_read & TID_OPFN_MAX_READ_MASK) <<
153 TID_OPFN_MAX_READ_SHIFT) |
154 (((u64)p->max_write & TID_OPFN_MAX_WRITE_MASK) <<
155 TID_OPFN_MAX_WRITE_SHIFT);
158 static void tid_rdma_opfn_decode(struct tid_rdma_params *p, u64 data)
160 p->max_len = (((data >> TID_OPFN_MAX_LEN_SHIFT) &
161 TID_OPFN_MAX_LEN_MASK) + 1) << PAGE_SHIFT;
162 p->jkey = (data >> TID_OPFN_JKEY_SHIFT) & TID_OPFN_JKEY_MASK;
163 p->max_write = (data >> TID_OPFN_MAX_WRITE_SHIFT) &
164 TID_OPFN_MAX_WRITE_MASK;
165 p->max_read = (data >> TID_OPFN_MAX_READ_SHIFT) &
166 TID_OPFN_MAX_READ_MASK;
168 ((((data >> TID_OPFN_QP_KDETH_SHIFT) & TID_OPFN_QP_KDETH_MASK)
170 ((data >> TID_OPFN_QP_CTXT_SHIFT) & TID_OPFN_QP_CTXT_MASK));
171 p->urg = (data >> TID_OPFN_URG_SHIFT) & TID_OPFN_URG_MASK;
172 p->timeout = (data >> TID_OPFN_TIMEOUT_SHIFT) & TID_OPFN_TIMEOUT_MASK;
175 void tid_rdma_opfn_init(struct rvt_qp *qp, struct tid_rdma_params *p)
177 struct hfi1_qp_priv *priv = qp->priv;
179 p->qp = (kdeth_qp << 16) | priv->rcd->ctxt;
180 p->max_len = TID_RDMA_MAX_SEGMENT_SIZE;
181 p->jkey = priv->rcd->jkey;
182 p->max_read = TID_RDMA_MAX_READ_SEGS_PER_REQ;
183 p->max_write = TID_RDMA_MAX_WRITE_SEGS_PER_REQ;
184 p->timeout = qp->timeout;
185 p->urg = is_urg_masked(priv->rcd);
188 bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data)
190 struct hfi1_qp_priv *priv = qp->priv;
192 *data = tid_rdma_opfn_encode(&priv->tid_rdma.local);
196 bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data)
198 struct hfi1_qp_priv *priv = qp->priv;
199 struct tid_rdma_params *remote, *old;
202 old = rcu_dereference_protected(priv->tid_rdma.remote,
203 lockdep_is_held(&priv->opfn.lock));
206 * If data passed in is zero, return true so as not to continue the
207 * negotiation process
209 if (!data || !HFI1_CAP_IS_KSET(TID_RDMA))
212 * If kzalloc fails, return false. This will result in:
213 * * at the requester a new OPFN request being generated to retry
215 * * at the responder, 0 being returned to the requester so as to
216 * disable TID RDMA at both the requester and the responder
218 remote = kzalloc(sizeof(*remote), GFP_ATOMIC);
224 tid_rdma_opfn_decode(remote, data);
225 priv->tid_timer_timeout_jiffies =
226 usecs_to_jiffies((((4096UL * (1UL << remote->timeout)) /
228 trace_hfi1_opfn_param(qp, 0, &priv->tid_rdma.local);
229 trace_hfi1_opfn_param(qp, 1, remote);
230 rcu_assign_pointer(priv->tid_rdma.remote, remote);
232 * A TID RDMA READ request's segment size is not equal to
233 * remote->max_len only when the request's data length is smaller
234 * than remote->max_len. In that case, there will be only one segment.
235 * Therefore, when priv->pkts_ps is used to calculate req->cur_seg
236 * during retry, it will lead to req->cur_seg = 0, which is exactly
239 priv->pkts_ps = (u16)rvt_div_mtu(qp, remote->max_len);
240 priv->timeout_shift = ilog2(priv->pkts_ps - 1) + 1;
243 RCU_INIT_POINTER(priv->tid_rdma.remote, NULL);
244 priv->timeout_shift = 0;
247 kfree_rcu(old, rcu_head);
251 bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data)
255 ret = tid_rdma_conn_reply(qp, *data);
258 * If tid_rdma_conn_reply() returns error, set *data as 0 to indicate
259 * TID RDMA could not be enabled. This will result in TID RDMA being
260 * disabled at the requester too.
263 (void)tid_rdma_conn_req(qp, data);
267 void tid_rdma_conn_error(struct rvt_qp *qp)
269 struct hfi1_qp_priv *priv = qp->priv;
270 struct tid_rdma_params *old;
272 old = rcu_dereference_protected(priv->tid_rdma.remote,
273 lockdep_is_held(&priv->opfn.lock));
274 RCU_INIT_POINTER(priv->tid_rdma.remote, NULL);
276 kfree_rcu(old, rcu_head);
279 /* This is called at context initialization time */
280 int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit)
285 BUILD_BUG_ON(TID_RDMA_JKEY < HFI1_KERNEL_MIN_JKEY);
286 BUILD_BUG_ON(TID_RDMA_JKEY > HFI1_KERNEL_MAX_JKEY);
287 rcd->jkey = TID_RDMA_JKEY;
288 hfi1_set_ctxt_jkey(rcd->dd, rcd, rcd->jkey);
289 return hfi1_alloc_ctxt_rcv_groups(rcd);
293 * qp_to_rcd - determine the receive context used by a qp
296 * This routine returns the receive context associated
299 * Returns the context.
301 static struct hfi1_ctxtdata *qp_to_rcd(struct rvt_dev_info *rdi,
304 struct hfi1_ibdev *verbs_dev = container_of(rdi,
307 struct hfi1_devdata *dd = container_of(verbs_dev,
312 if (qp->ibqp.qp_num == 0)
315 ctxt = hfi1_get_qp_map(dd, qp->ibqp.qp_num >> dd->qos_shift);
316 return dd->rcd[ctxt];
319 int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp,
320 struct ib_qp_init_attr *init_attr)
322 struct hfi1_qp_priv *qpriv = qp->priv;
325 qpriv->rcd = qp_to_rcd(rdi, qp);
327 spin_lock_init(&qpriv->opfn.lock);
328 INIT_WORK(&qpriv->opfn.opfn_work, opfn_send_conn_request);
329 INIT_WORK(&qpriv->tid_rdma.trigger_work, tid_rdma_trigger_resume);
330 qpriv->flow_state.psn = 0;
331 qpriv->flow_state.index = RXE_NUM_TID_FLOWS;
332 qpriv->flow_state.last_index = RXE_NUM_TID_FLOWS;
333 qpriv->flow_state.generation = KERN_GENERATION_RESERVED;
334 qpriv->s_state = TID_OP(WRITE_RESP);
335 qpriv->s_tid_cur = HFI1_QP_WQE_INVALID;
336 qpriv->s_tid_head = HFI1_QP_WQE_INVALID;
337 qpriv->s_tid_tail = HFI1_QP_WQE_INVALID;
338 qpriv->rnr_nak_state = TID_RNR_NAK_INIT;
339 qpriv->r_tid_head = HFI1_QP_WQE_INVALID;
340 qpriv->r_tid_tail = HFI1_QP_WQE_INVALID;
341 qpriv->r_tid_ack = HFI1_QP_WQE_INVALID;
342 qpriv->r_tid_alloc = HFI1_QP_WQE_INVALID;
343 atomic_set(&qpriv->n_requests, 0);
344 atomic_set(&qpriv->n_tid_requests, 0);
345 timer_setup(&qpriv->s_tid_timer, hfi1_tid_timeout, 0);
346 timer_setup(&qpriv->s_tid_retry_timer, hfi1_tid_retry_timeout, 0);
347 INIT_LIST_HEAD(&qpriv->tid_wait);
349 if (init_attr->qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) {
350 struct hfi1_devdata *dd = qpriv->rcd->dd;
352 qpriv->pages = kzalloc_node(TID_RDMA_MAX_PAGES *
353 sizeof(*qpriv->pages),
354 GFP_KERNEL, dd->node);
357 for (i = 0; i < qp->s_size; i++) {
358 struct hfi1_swqe_priv *priv;
359 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, i);
361 priv = kzalloc_node(sizeof(*priv), GFP_KERNEL,
366 hfi1_init_trdma_req(qp, &priv->tid_req);
367 priv->tid_req.e.swqe = wqe;
370 for (i = 0; i < rvt_max_atomic(rdi); i++) {
371 struct hfi1_ack_priv *priv;
373 priv = kzalloc_node(sizeof(*priv), GFP_KERNEL,
378 hfi1_init_trdma_req(qp, &priv->tid_req);
379 priv->tid_req.e.ack = &qp->s_ack_queue[i];
381 ret = hfi1_kern_exp_rcv_alloc_flows(&priv->tid_req,
387 qp->s_ack_queue[i].priv = priv;
394 void hfi1_qp_priv_tid_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
396 struct hfi1_qp_priv *qpriv = qp->priv;
397 struct rvt_swqe *wqe;
400 if (qp->ibqp.qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) {
401 for (i = 0; i < qp->s_size; i++) {
402 wqe = rvt_get_swqe_ptr(qp, i);
406 for (i = 0; i < rvt_max_atomic(rdi); i++) {
407 struct hfi1_ack_priv *priv = qp->s_ack_queue[i].priv;
410 hfi1_kern_exp_rcv_free_flows(&priv->tid_req);
412 qp->s_ack_queue[i].priv = NULL;
414 cancel_work_sync(&qpriv->opfn.opfn_work);
420 /* Flow and tid waiter functions */
424 * There are two locks involved with the queuing
425 * routines: the qp s_lock and the exp_lock.
427 * Since the tid space allocation is called from
428 * the send engine, the qp s_lock is already held.
430 * The allocation routines will get the exp_lock.
432 * The first_qp() call is provided to allow the head of
433 * the rcd wait queue to be fetched under the exp_lock and
434 * followed by a drop of the exp_lock.
436 * Any qp in the wait list will have the qp reference count held
437 * to hold the qp in memory.
441 * return head of rcd wait list
443 * Must hold the exp_lock.
445 * Get a reference to the QP to hold the QP in memory.
447 * The caller must release the reference when the local
448 * is no longer being used.
450 static struct rvt_qp *first_qp(struct hfi1_ctxtdata *rcd,
451 struct tid_queue *queue)
452 __must_hold(&rcd->exp_lock)
454 struct hfi1_qp_priv *priv;
456 lockdep_assert_held(&rcd->exp_lock);
457 priv = list_first_entry_or_null(&queue->queue_head,
462 rvt_get_qp(priv->owner);
467 * kernel_tid_waiters - determine rcd wait
468 * @rcd: the receive context
469 * @qp: the head of the qp being processed
471 * This routine will return false IFF
472 * the list is NULL or the head of the
473 * list is the indicated qp.
475 * Must hold the qp s_lock and the exp_lock.
478 * false if either of the conditions below are statisfied:
479 * 1. The list is empty or
480 * 2. The indicated qp is at the head of the list and the
481 * HFI1_S_WAIT_TID_SPACE bit is set in qp->s_flags.
482 * true is returned otherwise.
484 static bool kernel_tid_waiters(struct hfi1_ctxtdata *rcd,
485 struct tid_queue *queue, struct rvt_qp *qp)
486 __must_hold(&rcd->exp_lock) __must_hold(&qp->s_lock)
491 lockdep_assert_held(&qp->s_lock);
492 lockdep_assert_held(&rcd->exp_lock);
493 fqp = first_qp(rcd, queue);
494 if (!fqp || (fqp == qp && (qp->s_flags & HFI1_S_WAIT_TID_SPACE)))
501 * dequeue_tid_waiter - dequeue the qp from the list
502 * @qp - the qp to remove the wait list
504 * This routine removes the indicated qp from the
505 * wait list if it is there.
507 * This should be done after the hardware flow and
508 * tid array resources have been allocated.
510 * Must hold the qp s_lock and the rcd exp_lock.
512 * It assumes the s_lock to protect the s_flags
513 * field and to reliably test the HFI1_S_WAIT_TID_SPACE flag.
515 static void dequeue_tid_waiter(struct hfi1_ctxtdata *rcd,
516 struct tid_queue *queue, struct rvt_qp *qp)
517 __must_hold(&rcd->exp_lock) __must_hold(&qp->s_lock)
519 struct hfi1_qp_priv *priv = qp->priv;
521 lockdep_assert_held(&qp->s_lock);
522 lockdep_assert_held(&rcd->exp_lock);
523 if (list_empty(&priv->tid_wait))
525 list_del_init(&priv->tid_wait);
526 qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE;
532 * queue_qp_for_tid_wait - suspend QP on tid space
533 * @rcd: the receive context
536 * The qp is inserted at the tail of the rcd
537 * wait queue and the HFI1_S_WAIT_TID_SPACE s_flag is set.
539 * Must hold the qp s_lock and the exp_lock.
541 static void queue_qp_for_tid_wait(struct hfi1_ctxtdata *rcd,
542 struct tid_queue *queue, struct rvt_qp *qp)
543 __must_hold(&rcd->exp_lock) __must_hold(&qp->s_lock)
545 struct hfi1_qp_priv *priv = qp->priv;
547 lockdep_assert_held(&qp->s_lock);
548 lockdep_assert_held(&rcd->exp_lock);
549 if (list_empty(&priv->tid_wait)) {
550 qp->s_flags |= HFI1_S_WAIT_TID_SPACE;
551 list_add_tail(&priv->tid_wait, &queue->queue_head);
552 priv->tid_enqueue = ++queue->enqueue;
553 rcd->dd->verbs_dev.n_tidwait++;
554 trace_hfi1_qpsleep(qp, HFI1_S_WAIT_TID_SPACE);
560 * __trigger_tid_waiter - trigger tid waiter
563 * This is a private entrance to schedule the qp
564 * assuming the caller is holding the qp->s_lock.
566 static void __trigger_tid_waiter(struct rvt_qp *qp)
567 __must_hold(&qp->s_lock)
569 lockdep_assert_held(&qp->s_lock);
570 if (!(qp->s_flags & HFI1_S_WAIT_TID_SPACE))
572 trace_hfi1_qpwakeup(qp, HFI1_S_WAIT_TID_SPACE);
573 hfi1_schedule_send(qp);
577 * tid_rdma_schedule_tid_wakeup - schedule wakeup for a qp
580 * trigger a schedule or a waiting qp in a deadlock
581 * safe manner. The qp reference is held prior
582 * to this call via first_qp().
584 * If the qp trigger was already scheduled (!rval)
585 * the the reference is dropped, otherwise the resume
586 * or the destroy cancel will dispatch the reference.
588 static void tid_rdma_schedule_tid_wakeup(struct rvt_qp *qp)
590 struct hfi1_qp_priv *priv;
591 struct hfi1_ibport *ibp;
592 struct hfi1_pportdata *ppd;
593 struct hfi1_devdata *dd;
600 ibp = to_iport(qp->ibqp.device, qp->port_num);
601 ppd = ppd_from_ibp(ibp);
602 dd = dd_from_ibdev(qp->ibqp.device);
604 rval = queue_work_on(priv->s_sde ?
606 cpumask_first(cpumask_of_node(dd->node)),
608 &priv->tid_rdma.trigger_work);
614 * tid_rdma_trigger_resume - field a trigger work request
615 * @work - the work item
617 * Complete the off qp trigger processing by directly
618 * calling the progress routine.
620 static void tid_rdma_trigger_resume(struct work_struct *work)
622 struct tid_rdma_qp_params *tr;
623 struct hfi1_qp_priv *priv;
626 tr = container_of(work, struct tid_rdma_qp_params, trigger_work);
627 priv = container_of(tr, struct hfi1_qp_priv, tid_rdma);
629 spin_lock_irq(&qp->s_lock);
630 if (qp->s_flags & HFI1_S_WAIT_TID_SPACE) {
631 spin_unlock_irq(&qp->s_lock);
632 hfi1_do_send(priv->owner, true);
634 spin_unlock_irq(&qp->s_lock);
640 * tid_rdma_flush_wait - unwind any tid space wait
642 * This is called when resetting a qp to
643 * allow a destroy or reset to get rid
644 * of any tid space linkage and reference counts.
646 static void _tid_rdma_flush_wait(struct rvt_qp *qp, struct tid_queue *queue)
647 __must_hold(&qp->s_lock)
649 struct hfi1_qp_priv *priv;
653 lockdep_assert_held(&qp->s_lock);
655 qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE;
656 spin_lock(&priv->rcd->exp_lock);
657 if (!list_empty(&priv->tid_wait)) {
658 list_del_init(&priv->tid_wait);
659 qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE;
663 spin_unlock(&priv->rcd->exp_lock);
666 void hfi1_tid_rdma_flush_wait(struct rvt_qp *qp)
667 __must_hold(&qp->s_lock)
669 struct hfi1_qp_priv *priv = qp->priv;
671 _tid_rdma_flush_wait(qp, &priv->rcd->flow_queue);
672 _tid_rdma_flush_wait(qp, &priv->rcd->rarr_queue);
677 * kern_reserve_flow - allocate a hardware flow
678 * @rcd - the context to use for allocation
679 * @last - the index of the preferred flow. Use RXE_NUM_TID_FLOWS to
680 * signify "don't care".
682 * Use a bit mask based allocation to reserve a hardware
683 * flow for use in receiving KDETH data packets. If a preferred flow is
684 * specified the function will attempt to reserve that flow again, if
687 * The exp_lock must be held.
690 * On success: a value postive value between 0 and RXE_NUM_TID_FLOWS - 1
691 * On failure: -EAGAIN
693 static int kern_reserve_flow(struct hfi1_ctxtdata *rcd, int last)
694 __must_hold(&rcd->exp_lock)
698 /* Attempt to reserve the preferred flow index */
699 if (last >= 0 && last < RXE_NUM_TID_FLOWS &&
700 !test_and_set_bit(last, &rcd->flow_mask))
703 nr = ffz(rcd->flow_mask);
704 BUILD_BUG_ON(RXE_NUM_TID_FLOWS >=
705 (sizeof(rcd->flow_mask) * BITS_PER_BYTE));
706 if (nr > (RXE_NUM_TID_FLOWS - 1))
708 set_bit(nr, &rcd->flow_mask);
712 static void kern_set_hw_flow(struct hfi1_ctxtdata *rcd, u32 generation,
717 reg = ((u64)generation << HFI1_KDETH_BTH_SEQ_SHIFT) |
718 RCV_TID_FLOW_TABLE_CTRL_FLOW_VALID_SMASK |
719 RCV_TID_FLOW_TABLE_CTRL_KEEP_AFTER_SEQ_ERR_SMASK |
720 RCV_TID_FLOW_TABLE_CTRL_KEEP_ON_GEN_ERR_SMASK |
721 RCV_TID_FLOW_TABLE_STATUS_SEQ_MISMATCH_SMASK |
722 RCV_TID_FLOW_TABLE_STATUS_GEN_MISMATCH_SMASK;
724 if (generation != KERN_GENERATION_RESERVED)
725 reg |= RCV_TID_FLOW_TABLE_CTRL_HDR_SUPP_EN_SMASK;
727 write_uctxt_csr(rcd->dd, rcd->ctxt,
728 RCV_TID_FLOW_TABLE + 8 * flow_idx, reg);
731 static u32 kern_setup_hw_flow(struct hfi1_ctxtdata *rcd, u32 flow_idx)
732 __must_hold(&rcd->exp_lock)
734 u32 generation = rcd->flows[flow_idx].generation;
736 kern_set_hw_flow(rcd, generation, flow_idx);
740 static u32 kern_flow_generation_next(u32 gen)
742 u32 generation = mask_generation(gen + 1);
744 if (generation == KERN_GENERATION_RESERVED)
745 generation = mask_generation(generation + 1);
749 static void kern_clear_hw_flow(struct hfi1_ctxtdata *rcd, u32 flow_idx)
750 __must_hold(&rcd->exp_lock)
752 rcd->flows[flow_idx].generation =
753 kern_flow_generation_next(rcd->flows[flow_idx].generation);
754 kern_set_hw_flow(rcd, KERN_GENERATION_RESERVED, flow_idx);
757 int hfi1_kern_setup_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp)
759 struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv;
760 struct tid_flow_state *fs = &qpriv->flow_state;
765 /* The QP already has an allocated flow */
766 if (fs->index != RXE_NUM_TID_FLOWS)
769 spin_lock_irqsave(&rcd->exp_lock, flags);
770 if (kernel_tid_waiters(rcd, &rcd->flow_queue, qp))
773 ret = kern_reserve_flow(rcd, fs->last_index);
777 fs->last_index = fs->index;
779 /* Generation received in a RESYNC overrides default flow generation */
780 if (fs->generation != KERN_GENERATION_RESERVED)
781 rcd->flows[fs->index].generation = fs->generation;
782 fs->generation = kern_setup_hw_flow(rcd, fs->index);
784 dequeue_tid_waiter(rcd, &rcd->flow_queue, qp);
785 /* get head before dropping lock */
786 fqp = first_qp(rcd, &rcd->flow_queue);
787 spin_unlock_irqrestore(&rcd->exp_lock, flags);
789 tid_rdma_schedule_tid_wakeup(fqp);
792 queue_qp_for_tid_wait(rcd, &rcd->flow_queue, qp);
793 spin_unlock_irqrestore(&rcd->exp_lock, flags);
797 void hfi1_kern_clear_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp)
799 struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv;
800 struct tid_flow_state *fs = &qpriv->flow_state;
804 if (fs->index >= RXE_NUM_TID_FLOWS)
806 spin_lock_irqsave(&rcd->exp_lock, flags);
807 kern_clear_hw_flow(rcd, fs->index);
808 clear_bit(fs->index, &rcd->flow_mask);
809 fs->index = RXE_NUM_TID_FLOWS;
811 fs->generation = KERN_GENERATION_RESERVED;
813 /* get head before dropping lock */
814 fqp = first_qp(rcd, &rcd->flow_queue);
815 spin_unlock_irqrestore(&rcd->exp_lock, flags);
818 __trigger_tid_waiter(fqp);
821 tid_rdma_schedule_tid_wakeup(fqp);
825 void hfi1_kern_init_ctxt_generations(struct hfi1_ctxtdata *rcd)
829 for (i = 0; i < RXE_NUM_TID_FLOWS; i++) {
830 rcd->flows[i].generation = mask_generation(prandom_u32());
831 kern_set_hw_flow(rcd, KERN_GENERATION_RESERVED, i);
835 /* TID allocation functions */
836 static u8 trdma_pset_order(struct tid_rdma_pageset *s)
840 return ilog2(count) + 1;
844 * tid_rdma_find_phys_blocks_4k - get groups base on mr info
845 * @npages - number of pages
846 * @pages - pointer to an array of page structs
847 * @list - page set array to return
849 * This routine returns the number of groups associated with
850 * the current sge information. This implementation is based
851 * on the expected receive find_phys_blocks() adjusted to
852 * use the MR information vs. the pfn.
855 * the number of RcvArray entries
857 static u32 tid_rdma_find_phys_blocks_4k(struct tid_rdma_flow *flow,
860 struct tid_rdma_pageset *list)
862 u32 pagecount, pageidx, setcount = 0, i;
863 void *vaddr, *this_vaddr;
869 * Look for sets of physically contiguous pages in the user buffer.
870 * This will allow us to optimize Expected RcvArray entry usage by
871 * using the bigger supported sizes.
873 vaddr = page_address(pages[0]);
874 trace_hfi1_tid_flow_page(flow->req->qp, flow, 0, 0, 0, vaddr);
875 for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
876 this_vaddr = i < npages ? page_address(pages[i]) : NULL;
877 trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 0, 0,
880 * If the vaddr's are not sequential, pages are not physically
883 if (this_vaddr != (vaddr + PAGE_SIZE)) {
885 * At this point we have to loop over the set of
886 * physically contiguous pages and break them down it
887 * sizes supported by the HW.
888 * There are two main constraints:
889 * 1. The max buffer size is MAX_EXPECTED_BUFFER.
890 * If the total set size is bigger than that
891 * program only a MAX_EXPECTED_BUFFER chunk.
892 * 2. The buffer size has to be a power of two. If
893 * it is not, round down to the closes power of
894 * 2 and program that size.
897 int maxpages = pagecount;
898 u32 bufsize = pagecount * PAGE_SIZE;
900 if (bufsize > MAX_EXPECTED_BUFFER)
902 MAX_EXPECTED_BUFFER >>
904 else if (!is_power_of_2(bufsize))
906 rounddown_pow_of_two(bufsize) >>
909 list[setcount].idx = pageidx;
910 list[setcount].count = maxpages;
911 trace_hfi1_tid_pageset(flow->req->qp, setcount,
913 list[setcount].count);
914 pagecount -= maxpages;
926 /* insure we always return an even number of sets */
928 list[setcount++].count = 0;
933 * tid_flush_pages - dump out pages into pagesets
934 * @list - list of pagesets
935 * @idx - pointer to current page index
936 * @pages - number of pages to dump
937 * @sets - current number of pagesset
939 * This routine flushes out accumuated pages.
941 * To insure an even number of sets the
942 * code may add a filler.
944 * This can happen with when pages is not
945 * a power of 2 or pages is a power of 2
946 * less than the maximum pages.
949 * The new number of sets
952 static u32 tid_flush_pages(struct tid_rdma_pageset *list,
953 u32 *idx, u32 pages, u32 sets)
956 u32 maxpages = pages;
958 if (maxpages > MAX_EXPECTED_PAGES)
959 maxpages = MAX_EXPECTED_PAGES;
960 else if (!is_power_of_2(maxpages))
961 maxpages = rounddown_pow_of_two(maxpages);
962 list[sets].idx = *idx;
963 list[sets++].count = maxpages;
967 /* might need a filler */
969 list[sets++].count = 0;
974 * tid_rdma_find_phys_blocks_8k - get groups base on mr info
975 * @pages - pointer to an array of page structs
976 * @npages - number of pages
977 * @list - page set array to return
979 * This routine parses an array of pages to compute pagesets
980 * in an 8k compatible way.
982 * pages are tested two at a time, i, i + 1 for contiguous
983 * pages and i - 1 and i contiguous pages.
985 * If any condition is false, any accumlated pages are flushed and
986 * v0,v1 are emitted as separate PAGE_SIZE pagesets
988 * Otherwise, the current 8k is totaled for a future flush.
991 * The number of pagesets
992 * list set with the returned number of pagesets
995 static u32 tid_rdma_find_phys_blocks_8k(struct tid_rdma_flow *flow,
998 struct tid_rdma_pageset *list)
1000 u32 idx, sets = 0, i;
1002 void *v0, *v1, *vm1;
1006 for (idx = 0, i = 0, vm1 = NULL; i < npages; i += 2) {
1008 v0 = page_address(pages[i]);
1009 trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 1, 0, v0);
1010 v1 = i + 1 < npages ?
1011 page_address(pages[i + 1]) : NULL;
1012 trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 1, 1, v1);
1013 /* compare i, i + 1 vaddr */
1014 if (v1 != (v0 + PAGE_SIZE)) {
1015 /* flush out pages */
1016 sets = tid_flush_pages(list, &idx, pagecnt, sets);
1017 /* output v0,v1 as two pagesets */
1018 list[sets].idx = idx++;
1019 list[sets++].count = 1;
1021 list[sets].count = 1;
1022 list[sets++].idx = idx++;
1024 list[sets++].count = 0;
1030 /* i,i+1 consecutive, look at i-1,i */
1031 if (vm1 && v0 != (vm1 + PAGE_SIZE)) {
1032 /* flush out pages */
1033 sets = tid_flush_pages(list, &idx, pagecnt, sets);
1036 /* pages will always be a multiple of 8k */
1040 /* move to next pair */
1042 /* dump residual pages at end */
1043 sets = tid_flush_pages(list, &idx, npages - idx, sets);
1044 /* by design cannot be odd sets */
1050 * Find pages for one segment of a sge array represented by @ss. The function
1051 * does not check the sge, the sge must have been checked for alignment with a
1052 * prior call to hfi1_kern_trdma_ok. Other sge checking is done as part of
1053 * rvt_lkey_ok and rvt_rkey_ok. Also, the function only modifies the local sge
1054 * copy maintained in @ss->sge, the original sge is not modified.
1056 * Unlike IB RDMA WRITE, we can't decrement ss->num_sge here because we are not
1057 * releasing the MR reference count at the same time. Otherwise, we'll "leak"
1058 * references to the MR. This difference requires that we keep track of progress
1059 * into the sg_list. This is done by the cur_seg cursor in the tid_rdma_request
1062 static u32 kern_find_pages(struct tid_rdma_flow *flow,
1063 struct page **pages,
1064 struct rvt_sge_state *ss, bool *last)
1066 struct tid_rdma_request *req = flow->req;
1067 struct rvt_sge *sge = &ss->sge;
1068 u32 length = flow->req->seg_len;
1069 u32 len = PAGE_SIZE;
1072 while (length && req->isge < ss->num_sge) {
1073 pages[i++] = virt_to_page(sge->vaddr);
1077 sge->sge_length -= len;
1078 if (!sge->sge_length) {
1079 if (++req->isge < ss->num_sge)
1080 *sge = ss->sg_list[req->isge - 1];
1081 } else if (sge->length == 0 && sge->mr->lkey) {
1082 if (++sge->n >= RVT_SEGSZ) {
1086 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
1087 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
1092 flow->length = flow->req->seg_len - length;
1093 *last = req->isge == ss->num_sge ? false : true;
1097 static void dma_unmap_flow(struct tid_rdma_flow *flow)
1099 struct hfi1_devdata *dd;
1101 struct tid_rdma_pageset *pset;
1103 dd = flow->req->rcd->dd;
1104 for (i = 0, pset = &flow->pagesets[0]; i < flow->npagesets;
1106 if (pset->count && pset->addr) {
1107 dma_unmap_page(&dd->pcidev->dev,
1109 PAGE_SIZE * pset->count,
1116 static int dma_map_flow(struct tid_rdma_flow *flow, struct page **pages)
1119 struct hfi1_devdata *dd = flow->req->rcd->dd;
1120 struct tid_rdma_pageset *pset;
1122 for (i = 0, pset = &flow->pagesets[0]; i < flow->npagesets;
1125 pset->addr = dma_map_page(&dd->pcidev->dev,
1128 PAGE_SIZE * pset->count,
1131 if (dma_mapping_error(&dd->pcidev->dev, pset->addr)) {
1132 dma_unmap_flow(flow);
1141 static inline bool dma_mapped(struct tid_rdma_flow *flow)
1143 return !!flow->pagesets[0].mapped;
1147 * Get pages pointers and identify contiguous physical memory chunks for a
1148 * segment. All segments are of length flow->req->seg_len.
1150 static int kern_get_phys_blocks(struct tid_rdma_flow *flow,
1151 struct page **pages,
1152 struct rvt_sge_state *ss, bool *last)
1156 /* Reuse previously computed pagesets, if any */
1157 if (flow->npagesets) {
1158 trace_hfi1_tid_flow_alloc(flow->req->qp, flow->req->setup_head,
1160 if (!dma_mapped(flow))
1161 return dma_map_flow(flow, pages);
1165 npages = kern_find_pages(flow, pages, ss, last);
1167 if (flow->req->qp->pmtu == enum_to_mtu(OPA_MTU_4096))
1169 tid_rdma_find_phys_blocks_4k(flow, pages, npages,
1173 tid_rdma_find_phys_blocks_8k(flow, pages, npages,
1176 return dma_map_flow(flow, pages);
1179 static inline void kern_add_tid_node(struct tid_rdma_flow *flow,
1180 struct hfi1_ctxtdata *rcd, char *s,
1181 struct tid_group *grp, u8 cnt)
1183 struct kern_tid_node *node = &flow->tnode[flow->tnode_cnt++];
1185 WARN_ON_ONCE(flow->tnode_cnt >=
1186 (TID_RDMA_MAX_SEGMENT_SIZE >> PAGE_SHIFT));
1187 if (WARN_ON_ONCE(cnt & 1))
1189 "unexpected odd allocation cnt %u map 0x%x used %u",
1190 cnt, grp->map, grp->used);
1193 node->map = grp->map;
1195 trace_hfi1_tid_node_add(flow->req->qp, s, flow->tnode_cnt - 1,
1196 grp->base, grp->map, grp->used, cnt);
1200 * Try to allocate pageset_count TID's from TID groups for a context
1202 * This function allocates TID's without moving groups between lists or
1203 * modifying grp->map. This is done as follows, being cogizant of the lists
1204 * between which the TID groups will move:
1205 * 1. First allocate complete groups of 8 TID's since this is more efficient,
1206 * these groups will move from group->full without affecting used
1207 * 2. If more TID's are needed allocate from used (will move from used->full or
1209 * 3. If we still don't have the required number of TID's go back and look again
1210 * at a complete group (will move from group->used)
1212 static int kern_alloc_tids(struct tid_rdma_flow *flow)
1214 struct hfi1_ctxtdata *rcd = flow->req->rcd;
1215 struct hfi1_devdata *dd = rcd->dd;
1216 u32 ngroups, pageidx = 0;
1217 struct tid_group *group = NULL, *used;
1220 flow->tnode_cnt = 0;
1221 ngroups = flow->npagesets / dd->rcv_entries.group_size;
1225 /* First look at complete groups */
1226 list_for_each_entry(group, &rcd->tid_group_list.list, list) {
1227 kern_add_tid_node(flow, rcd, "complete groups", group,
1230 pageidx += group->size;
1235 if (pageidx >= flow->npagesets)
1239 /* Now look at partially used groups */
1240 list_for_each_entry(used, &rcd->tid_used_list.list, list) {
1241 use = min_t(u32, flow->npagesets - pageidx,
1242 used->size - used->used);
1243 kern_add_tid_node(flow, rcd, "used groups", used, use);
1246 if (pageidx >= flow->npagesets)
1251 * Look again at a complete group, continuing from where we left.
1252 * However, if we are at the head, we have reached the end of the
1253 * complete groups list from the first loop above
1255 if (group && &group->list == &rcd->tid_group_list.list)
1257 group = list_prepare_entry(group, &rcd->tid_group_list.list,
1259 if (list_is_last(&group->list, &rcd->tid_group_list.list))
1261 group = list_next_entry(group, list);
1262 use = min_t(u32, flow->npagesets - pageidx, group->size);
1263 kern_add_tid_node(flow, rcd, "complete continue", group, use);
1265 if (pageidx >= flow->npagesets)
1268 trace_hfi1_msg_alloc_tids(flow->req->qp, " insufficient tids: needed ",
1269 (u64)flow->npagesets);
1275 static void kern_program_rcv_group(struct tid_rdma_flow *flow, int grp_num,
1278 struct hfi1_ctxtdata *rcd = flow->req->rcd;
1279 struct hfi1_devdata *dd = rcd->dd;
1280 struct kern_tid_node *node = &flow->tnode[grp_num];
1281 struct tid_group *grp = node->grp;
1282 struct tid_rdma_pageset *pset;
1283 u32 pmtu_pg = flow->req->qp->pmtu >> PAGE_SHIFT;
1284 u32 rcventry, npages = 0, pair = 0, tidctrl;
1287 for (i = 0; i < grp->size; i++) {
1288 rcventry = grp->base + i;
1290 if (node->map & BIT(i) || cnt >= node->cnt) {
1291 rcv_array_wc_fill(dd, rcventry);
1294 pset = &flow->pagesets[(*pset_idx)++];
1296 hfi1_put_tid(dd, rcventry, PT_EXPECTED,
1297 pset->addr, trdma_pset_order(pset));
1299 hfi1_put_tid(dd, rcventry, PT_INVALID, 0, 0);
1301 npages += pset->count;
1303 rcventry -= rcd->expected_base;
1304 tidctrl = pair ? 0x3 : rcventry & 0x1 ? 0x2 : 0x1;
1306 * A single TID entry will be used to use a rcvarr pair (with
1307 * tidctrl 0x3), if ALL these are true (a) the bit pos is even
1308 * (b) the group map shows current and the next bits as free
1309 * indicating two consecutive rcvarry entries are available (c)
1310 * we actually need 2 more entries
1312 pair = !(i & 0x1) && !((node->map >> i) & 0x3) &&
1313 node->cnt >= cnt + 2;
1317 flow->tid_entry[flow->tidcnt++] =
1318 EXP_TID_SET(IDX, rcventry >> 1) |
1319 EXP_TID_SET(CTRL, tidctrl) |
1320 EXP_TID_SET(LEN, npages);
1321 trace_hfi1_tid_entry_alloc(/* entry */
1322 flow->req->qp, flow->tidcnt - 1,
1323 flow->tid_entry[flow->tidcnt - 1]);
1325 /* Efficient DIV_ROUND_UP(npages, pmtu_pg) */
1326 flow->npkts += (npages + pmtu_pg - 1) >> ilog2(pmtu_pg);
1330 if (grp->used == grp->size - 1)
1331 tid_group_move(grp, &rcd->tid_used_list,
1332 &rcd->tid_full_list);
1333 else if (!grp->used)
1334 tid_group_move(grp, &rcd->tid_group_list,
1335 &rcd->tid_used_list);
1343 static void kern_unprogram_rcv_group(struct tid_rdma_flow *flow, int grp_num)
1345 struct hfi1_ctxtdata *rcd = flow->req->rcd;
1346 struct hfi1_devdata *dd = rcd->dd;
1347 struct kern_tid_node *node = &flow->tnode[grp_num];
1348 struct tid_group *grp = node->grp;
1352 for (i = 0; i < grp->size; i++) {
1353 rcventry = grp->base + i;
1355 if (node->map & BIT(i) || cnt >= node->cnt) {
1356 rcv_array_wc_fill(dd, rcventry);
1360 hfi1_put_tid(dd, rcventry, PT_INVALID, 0, 0);
1363 grp->map &= ~BIT(i);
1366 if (grp->used == grp->size - 1)
1367 tid_group_move(grp, &rcd->tid_full_list,
1368 &rcd->tid_used_list);
1369 else if (!grp->used)
1370 tid_group_move(grp, &rcd->tid_used_list,
1371 &rcd->tid_group_list);
1373 if (WARN_ON_ONCE(cnt & 1)) {
1374 struct hfi1_ctxtdata *rcd = flow->req->rcd;
1375 struct hfi1_devdata *dd = rcd->dd;
1377 dd_dev_err(dd, "unexpected odd free cnt %u map 0x%x used %u",
1378 cnt, grp->map, grp->used);
1382 static void kern_program_rcvarray(struct tid_rdma_flow *flow)
1389 for (i = 0; i < flow->tnode_cnt; i++)
1390 kern_program_rcv_group(flow, i, &pset_idx);
1391 trace_hfi1_tid_flow_alloc(flow->req->qp, flow->req->setup_head, flow);
1395 * hfi1_kern_exp_rcv_setup() - setup TID's and flow for one segment of a
1398 * @req: TID RDMA request for which the segment/flow is being set up
1399 * @ss: sge state, maintains state across successive segments of a sge
1400 * @last: set to true after the last sge segment has been processed
1403 * (1) finds a free flow entry in the flow circular buffer
1404 * (2) finds pages and continuous physical chunks constituing one segment
1406 * (3) allocates TID group entries for those chunks
1407 * (4) programs rcvarray entries in the hardware corresponding to those
1409 * (5) computes a tidarray with formatted TID entries which can be sent
1411 * (6) Reserves and programs HW flows.
1412 * (7) It also manages queing the QP when TID/flow resources are not
1415 * @req points to struct tid_rdma_request of which the segments are a part. The
1416 * function uses qp, rcd and seg_len members of @req. In the absence of errors,
1417 * req->flow_idx is the index of the flow which has been prepared in this
1418 * invocation of function call. With flow = &req->flows[req->flow_idx],
1419 * flow->tid_entry contains the TID array which the sender can use for TID RDMA
1420 * sends and flow->npkts contains number of packets required to send the
1423 * hfi1_check_sge_align should be called prior to calling this function and if
1424 * it signals error TID RDMA cannot be used for this sge and this function
1425 * should not be called.
1427 * For the queuing, caller must hold the flow->req->qp s_lock from the send
1428 * engine and the function will procure the exp_lock.
1431 * The function returns -EAGAIN if sufficient number of TID/flow resources to
1432 * map the segment could not be allocated. In this case the function should be
1433 * called again with previous arguments to retry the TID allocation. There are
1434 * no other error returns. The function returns 0 on success.
1436 int hfi1_kern_exp_rcv_setup(struct tid_rdma_request *req,
1437 struct rvt_sge_state *ss, bool *last)
1438 __must_hold(&req->qp->s_lock)
1440 struct tid_rdma_flow *flow = &req->flows[req->setup_head];
1441 struct hfi1_ctxtdata *rcd = req->rcd;
1442 struct hfi1_qp_priv *qpriv = req->qp->priv;
1443 unsigned long flags;
1445 u16 clear_tail = req->clear_tail;
1447 lockdep_assert_held(&req->qp->s_lock);
1449 * We return error if either (a) we don't have space in the flow
1450 * circular buffer, or (b) we already have max entries in the buffer.
1451 * Max entries depend on the type of request we are processing and the
1452 * negotiated TID RDMA parameters.
1454 if (!CIRC_SPACE(req->setup_head, clear_tail, MAX_FLOWS) ||
1455 CIRC_CNT(req->setup_head, clear_tail, MAX_FLOWS) >=
1460 * Get pages, identify contiguous physical memory chunks for the segment
1461 * If we can not determine a DMA address mapping we will treat it just
1462 * like if we ran out of space above.
1464 if (kern_get_phys_blocks(flow, qpriv->pages, ss, last)) {
1465 hfi1_wait_kmem(flow->req->qp);
1469 spin_lock_irqsave(&rcd->exp_lock, flags);
1470 if (kernel_tid_waiters(rcd, &rcd->rarr_queue, flow->req->qp))
1474 * At this point we know the number of pagesets and hence the number of
1475 * TID's to map the segment. Allocate the TID's from the TID groups. If
1476 * we cannot allocate the required number we exit and try again later
1478 if (kern_alloc_tids(flow))
1481 * Finally program the TID entries with the pagesets, compute the
1482 * tidarray and enable the HW flow
1484 kern_program_rcvarray(flow);
1487 * Setup the flow state with relevant information.
1488 * This information is used for tracking the sequence of data packets
1490 * The flow is setup here as this is the most accurate time and place
1491 * to do so. Doing at a later time runs the risk of the flow data in
1492 * qpriv getting out of sync.
1494 memset(&flow->flow_state, 0x0, sizeof(flow->flow_state));
1495 flow->idx = qpriv->flow_state.index;
1496 flow->flow_state.generation = qpriv->flow_state.generation;
1497 flow->flow_state.spsn = qpriv->flow_state.psn;
1498 flow->flow_state.lpsn = flow->flow_state.spsn + flow->npkts - 1;
1499 flow->flow_state.r_next_psn =
1500 full_flow_psn(flow, flow->flow_state.spsn);
1501 qpriv->flow_state.psn += flow->npkts;
1503 dequeue_tid_waiter(rcd, &rcd->rarr_queue, flow->req->qp);
1504 /* get head before dropping lock */
1505 fqp = first_qp(rcd, &rcd->rarr_queue);
1506 spin_unlock_irqrestore(&rcd->exp_lock, flags);
1507 tid_rdma_schedule_tid_wakeup(fqp);
1509 req->setup_head = (req->setup_head + 1) & (MAX_FLOWS - 1);
1512 queue_qp_for_tid_wait(rcd, &rcd->rarr_queue, flow->req->qp);
1513 spin_unlock_irqrestore(&rcd->exp_lock, flags);
1517 static void hfi1_tid_rdma_reset_flow(struct tid_rdma_flow *flow)
1519 flow->npagesets = 0;
1523 * This function is called after one segment has been successfully sent to
1524 * release the flow and TID HW/SW resources for that segment. The segments for a
1525 * TID RDMA request are setup and cleared in FIFO order which is managed using a
1528 int hfi1_kern_exp_rcv_clear(struct tid_rdma_request *req)
1529 __must_hold(&req->qp->s_lock)
1531 struct tid_rdma_flow *flow = &req->flows[req->clear_tail];
1532 struct hfi1_ctxtdata *rcd = req->rcd;
1533 unsigned long flags;
1537 lockdep_assert_held(&req->qp->s_lock);
1538 /* Exit if we have nothing in the flow circular buffer */
1539 if (!CIRC_CNT(req->setup_head, req->clear_tail, MAX_FLOWS))
1542 spin_lock_irqsave(&rcd->exp_lock, flags);
1544 for (i = 0; i < flow->tnode_cnt; i++)
1545 kern_unprogram_rcv_group(flow, i);
1546 /* To prevent double unprogramming */
1547 flow->tnode_cnt = 0;
1548 /* get head before dropping lock */
1549 fqp = first_qp(rcd, &rcd->rarr_queue);
1550 spin_unlock_irqrestore(&rcd->exp_lock, flags);
1552 dma_unmap_flow(flow);
1554 hfi1_tid_rdma_reset_flow(flow);
1555 req->clear_tail = (req->clear_tail + 1) & (MAX_FLOWS - 1);
1557 if (fqp == req->qp) {
1558 __trigger_tid_waiter(fqp);
1561 tid_rdma_schedule_tid_wakeup(fqp);
1568 * This function is called to release all the tid entries for
1571 void hfi1_kern_exp_rcv_clear_all(struct tid_rdma_request *req)
1572 __must_hold(&req->qp->s_lock)
1574 /* Use memory barrier for proper ordering */
1575 while (CIRC_CNT(req->setup_head, req->clear_tail, MAX_FLOWS)) {
1576 if (hfi1_kern_exp_rcv_clear(req))
1582 * hfi1_kern_exp_rcv_free_flows - free priviously allocated flow information
1583 * @req - the tid rdma request to be cleaned
1585 static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req)
1592 * __trdma_clean_swqe - clean up for large sized QPs
1593 * @qp: the queue patch
1594 * @wqe: the send wqe
1596 void __trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe)
1598 struct hfi1_swqe_priv *p = wqe->priv;
1600 hfi1_kern_exp_rcv_free_flows(&p->tid_req);
1604 * This can be called at QP create time or in the data path.
1606 static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req,
1609 struct tid_rdma_flow *flows;
1612 if (likely(req->flows))
1614 flows = kmalloc_node(MAX_FLOWS * sizeof(*flows), gfp,
1619 for (i = 0; i < MAX_FLOWS; i++) {
1621 flows[i].npagesets = 0;
1622 flows[i].pagesets[0].mapped = 0;
1628 static void hfi1_init_trdma_req(struct rvt_qp *qp,
1629 struct tid_rdma_request *req)
1631 struct hfi1_qp_priv *qpriv = qp->priv;
1634 * Initialize various TID RDMA request variables.
1635 * These variables are "static", which is why they
1636 * can be pre-initialized here before the WRs has
1637 * even been submitted.
1638 * However, non-NULL values for these variables do not
1639 * imply that this WQE has been enabled for TID RDMA.
1640 * Drivers should check the WQE's opcode to determine
1641 * if a request is a TID RDMA one or not.
1644 req->rcd = qpriv->rcd;
1647 u64 hfi1_access_sw_tid_wait(const struct cntr_entry *entry,
1648 void *context, int vl, int mode, u64 data)
1650 struct hfi1_devdata *dd = context;
1652 return dd->verbs_dev.n_tidwait;
1655 static struct tid_rdma_flow *find_flow_ib(struct tid_rdma_request *req,
1659 struct tid_rdma_flow *flow;
1661 head = req->setup_head;
1662 tail = req->clear_tail;
1663 for ( ; CIRC_CNT(head, tail, MAX_FLOWS);
1664 tail = CIRC_NEXT(tail, MAX_FLOWS)) {
1665 flow = &req->flows[tail];
1666 if (cmp_psn(psn, flow->flow_state.ib_spsn) >= 0 &&
1667 cmp_psn(psn, flow->flow_state.ib_lpsn) <= 0) {
1676 static struct tid_rdma_flow *
1677 __find_flow_ranged(struct tid_rdma_request *req, u16 head, u16 tail,
1680 for ( ; CIRC_CNT(head, tail, MAX_FLOWS);
1681 tail = CIRC_NEXT(tail, MAX_FLOWS)) {
1682 struct tid_rdma_flow *flow = &req->flows[tail];
1685 spsn = full_flow_psn(flow, flow->flow_state.spsn);
1686 lpsn = full_flow_psn(flow, flow->flow_state.lpsn);
1688 if (cmp_psn(psn, spsn) >= 0 && cmp_psn(psn, lpsn) <= 0) {
1697 static struct tid_rdma_flow *find_flow(struct tid_rdma_request *req,
1700 return __find_flow_ranged(req, req->setup_head, req->clear_tail, psn,
1704 /* TID RDMA READ functions */
1705 u32 hfi1_build_tid_rdma_read_packet(struct rvt_swqe *wqe,
1706 struct ib_other_headers *ohdr, u32 *bth1,
1707 u32 *bth2, u32 *len)
1709 struct tid_rdma_request *req = wqe_to_tid_req(wqe);
1710 struct tid_rdma_flow *flow = &req->flows[req->flow_idx];
1711 struct rvt_qp *qp = req->qp;
1712 struct hfi1_qp_priv *qpriv = qp->priv;
1713 struct hfi1_swqe_priv *wpriv = wqe->priv;
1714 struct tid_rdma_read_req *rreq = &ohdr->u.tid_rdma.r_req;
1715 struct tid_rdma_params *remote;
1717 void *req_addr = NULL;
1719 /* This is the IB psn used to send the request */
1720 *bth2 = mask_psn(flow->flow_state.ib_spsn + flow->pkt);
1721 trace_hfi1_tid_flow_build_read_pkt(qp, req->flow_idx, flow);
1723 /* TID Entries for TID RDMA READ payload */
1724 req_addr = &flow->tid_entry[flow->tid_idx];
1725 req_len = sizeof(*flow->tid_entry) *
1726 (flow->tidcnt - flow->tid_idx);
1728 memset(&ohdr->u.tid_rdma.r_req, 0, sizeof(ohdr->u.tid_rdma.r_req));
1729 wpriv->ss.sge.vaddr = req_addr;
1730 wpriv->ss.sge.sge_length = req_len;
1731 wpriv->ss.sge.length = wpriv->ss.sge.sge_length;
1733 * We can safely zero these out. Since the first SGE covers the
1734 * entire packet, nothing else should even look at the MR.
1736 wpriv->ss.sge.mr = NULL;
1737 wpriv->ss.sge.m = 0;
1738 wpriv->ss.sge.n = 0;
1740 wpriv->ss.sg_list = NULL;
1741 wpriv->ss.total_len = wpriv->ss.sge.sge_length;
1742 wpriv->ss.num_sge = 1;
1744 /* Construct the TID RDMA READ REQ packet header */
1746 remote = rcu_dereference(qpriv->tid_rdma.remote);
1748 KDETH_RESET(rreq->kdeth0, KVER, 0x1);
1749 KDETH_RESET(rreq->kdeth1, JKEY, remote->jkey);
1750 rreq->reth.vaddr = cpu_to_be64(wqe->rdma_wr.remote_addr +
1751 req->cur_seg * req->seg_len + flow->sent);
1752 rreq->reth.rkey = cpu_to_be32(wqe->rdma_wr.rkey);
1753 rreq->reth.length = cpu_to_be32(*len);
1754 rreq->tid_flow_psn =
1755 cpu_to_be32((flow->flow_state.generation <<
1756 HFI1_KDETH_BTH_SEQ_SHIFT) |
1757 ((flow->flow_state.spsn + flow->pkt) &
1758 HFI1_KDETH_BTH_SEQ_MASK));
1760 cpu_to_be32(qpriv->tid_rdma.local.qp |
1761 ((flow->idx & TID_RDMA_DESTQP_FLOW_MASK) <<
1762 TID_RDMA_DESTQP_FLOW_SHIFT) |
1764 rreq->verbs_qp = cpu_to_be32(qp->remote_qpn);
1765 *bth1 &= ~RVT_QPN_MASK;
1766 *bth1 |= remote->qp;
1767 *bth2 |= IB_BTH_REQ_ACK;
1770 /* We are done with this segment */
1773 qp->s_state = TID_OP(READ_REQ);
1775 req->flow_idx = (req->flow_idx + 1) & (MAX_FLOWS - 1);
1776 qpriv->pending_tid_r_segs++;
1777 qp->s_num_rd_atomic++;
1779 /* Set the TID RDMA READ request payload size */
1782 return sizeof(ohdr->u.tid_rdma.r_req) / sizeof(u32);
1786 * @len: contains the data length to read upon entry and the read request
1787 * payload length upon exit.
1789 u32 hfi1_build_tid_rdma_read_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
1790 struct ib_other_headers *ohdr, u32 *bth1,
1791 u32 *bth2, u32 *len)
1792 __must_hold(&qp->s_lock)
1794 struct hfi1_qp_priv *qpriv = qp->priv;
1795 struct tid_rdma_request *req = wqe_to_tid_req(wqe);
1796 struct tid_rdma_flow *flow = NULL;
1800 u32 npkts = rvt_div_round_up_mtu(qp, *len);
1802 trace_hfi1_tid_req_build_read_req(qp, 0, wqe->wr.opcode, wqe->psn,
1805 * Check sync conditions. Make sure that there are no pending
1806 * segments before freeing the flow.
1809 if (req->state == TID_REQUEST_SYNC) {
1810 if (qpriv->pending_tid_r_segs)
1813 hfi1_kern_clear_hw_flow(req->rcd, qp);
1814 qpriv->s_flags &= ~HFI1_R_TID_SW_PSN;
1815 req->state = TID_REQUEST_ACTIVE;
1819 * If the request for this segment is resent, the tid resources should
1820 * have been allocated before. In this case, req->flow_idx should
1821 * fall behind req->setup_head.
1823 if (req->flow_idx == req->setup_head) {
1825 if (req->state == TID_REQUEST_RESEND) {
1827 * This is the first new segment for a request whose
1828 * earlier segments have been re-sent. We need to
1829 * set up the sge pointer correctly.
1831 restart_sge(&qp->s_sge, wqe, req->s_next_psn,
1834 req->state = TID_REQUEST_ACTIVE;
1838 * Check sync. The last PSN of each generation is reserved for
1841 if ((qpriv->flow_state.psn + npkts) > MAX_TID_FLOW_PSN - 1) {
1842 req->state = TID_REQUEST_SYNC;
1846 /* Allocate the flow if not yet */
1847 if (hfi1_kern_setup_hw_flow(qpriv->rcd, qp))
1851 * The following call will advance req->setup_head after
1852 * allocating the tid entries.
1854 if (hfi1_kern_exp_rcv_setup(req, &qp->s_sge, &last)) {
1855 req->state = TID_REQUEST_QUEUED;
1858 * We don't have resources for this segment. The QP has
1859 * already been queued.
1865 /* req->flow_idx should only be one slot behind req->setup_head */
1866 flow = &req->flows[req->flow_idx];
1871 /* Set the first and last IB PSN for the flow in use.*/
1872 flow->flow_state.ib_spsn = req->s_next_psn;
1873 flow->flow_state.ib_lpsn =
1874 flow->flow_state.ib_spsn + flow->npkts - 1;
1877 /* Calculate the next segment start psn.*/
1878 req->s_next_psn += flow->npkts;
1880 /* Build the packet header */
1881 hdwords = hfi1_build_tid_rdma_read_packet(wqe, ohdr, bth1, bth2, len);
1887 * Validate and accept the TID RDMA READ request parameters.
1888 * Return 0 if the request is accepted successfully;
1889 * Return 1 otherwise.
1891 static int tid_rdma_rcv_read_request(struct rvt_qp *qp,
1892 struct rvt_ack_entry *e,
1893 struct hfi1_packet *packet,
1894 struct ib_other_headers *ohdr,
1895 u32 bth0, u32 psn, u64 vaddr, u32 len)
1897 struct hfi1_qp_priv *qpriv = qp->priv;
1898 struct tid_rdma_request *req;
1899 struct tid_rdma_flow *flow;
1900 u32 flow_psn, i, tidlen = 0, pktlen, tlen;
1902 req = ack_to_tid_req(e);
1904 /* Validate the payload first */
1905 flow = &req->flows[req->setup_head];
1907 /* payload length = packet length - (header length + ICRC length) */
1908 pktlen = packet->tlen - (packet->hlen + 4);
1909 if (pktlen > sizeof(flow->tid_entry))
1911 memcpy(flow->tid_entry, packet->ebuf, pktlen);
1912 flow->tidcnt = pktlen / sizeof(*flow->tid_entry);
1915 * Walk the TID_ENTRY list to make sure we have enough space for a
1916 * complete segment. Also calculate the number of required packets.
1918 flow->npkts = rvt_div_round_up_mtu(qp, len);
1919 for (i = 0; i < flow->tidcnt; i++) {
1920 trace_hfi1_tid_entry_rcv_read_req(qp, i,
1921 flow->tid_entry[i]);
1922 tlen = EXP_TID_GET(flow->tid_entry[i], LEN);
1927 * For tid pair (tidctr == 3), the buffer size of the pair
1928 * should be the sum of the buffer size described by each
1929 * tid entry. However, only the first entry needs to be
1930 * specified in the request (see WFR HAS Section 8.5.7.1).
1934 if (tidlen * PAGE_SIZE < len)
1937 /* Empty the flow array */
1938 req->clear_tail = req->setup_head;
1941 flow->tid_offset = 0;
1943 flow->tid_qpn = be32_to_cpu(ohdr->u.tid_rdma.r_req.tid_flow_qp);
1944 flow->idx = (flow->tid_qpn >> TID_RDMA_DESTQP_FLOW_SHIFT) &
1945 TID_RDMA_DESTQP_FLOW_MASK;
1946 flow_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_req.tid_flow_psn));
1947 flow->flow_state.generation = flow_psn >> HFI1_KDETH_BTH_SEQ_SHIFT;
1948 flow->flow_state.spsn = flow_psn & HFI1_KDETH_BTH_SEQ_MASK;
1951 flow->flow_state.lpsn = flow->flow_state.spsn +
1953 flow->flow_state.ib_spsn = psn;
1954 flow->flow_state.ib_lpsn = flow->flow_state.ib_spsn + flow->npkts - 1;
1956 trace_hfi1_tid_flow_rcv_read_req(qp, req->setup_head, flow);
1957 /* Set the initial flow index to the current flow. */
1958 req->flow_idx = req->setup_head;
1960 /* advance circular buffer head */
1961 req->setup_head = (req->setup_head + 1) & (MAX_FLOWS - 1);
1964 * Compute last PSN for request.
1966 e->opcode = (bth0 >> 24) & 0xff;
1968 e->lpsn = psn + flow->npkts - 1;
1971 req->n_flows = qpriv->tid_rdma.local.max_read;
1972 req->state = TID_REQUEST_ACTIVE;
1977 req->seg_len = qpriv->tid_rdma.local.max_len;
1978 req->total_len = len;
1979 req->total_segs = 1;
1980 req->r_flow_psn = e->psn;
1982 trace_hfi1_tid_req_rcv_read_req(qp, 0, e->opcode, e->psn, e->lpsn,
1987 static int tid_rdma_rcv_error(struct hfi1_packet *packet,
1988 struct ib_other_headers *ohdr,
1989 struct rvt_qp *qp, u32 psn, int diff)
1991 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1992 struct hfi1_ctxtdata *rcd = ((struct hfi1_qp_priv *)qp->priv)->rcd;
1993 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
1994 struct hfi1_qp_priv *qpriv = qp->priv;
1995 struct rvt_ack_entry *e;
1996 struct tid_rdma_request *req;
1997 unsigned long flags;
2001 trace_hfi1_rsp_tid_rcv_error(qp, psn);
2002 trace_hfi1_tid_rdma_rcv_err(qp, 0, psn, diff);
2004 /* sequence error */
2005 if (!qp->r_nak_state) {
2006 ibp->rvp.n_rc_seqnak++;
2007 qp->r_nak_state = IB_NAK_PSN_ERROR;
2008 qp->r_ack_psn = qp->r_psn;
2009 rc_defered_ack(rcd, qp);
2014 ibp->rvp.n_rc_dupreq++;
2016 spin_lock_irqsave(&qp->s_lock, flags);
2017 e = find_prev_entry(qp, psn, &prev, NULL, &old_req);
2018 if (!e || (e->opcode != TID_OP(READ_REQ) &&
2019 e->opcode != TID_OP(WRITE_REQ)))
2022 req = ack_to_tid_req(e);
2023 req->r_flow_psn = psn;
2024 trace_hfi1_tid_req_rcv_err(qp, 0, e->opcode, e->psn, e->lpsn, req);
2025 if (e->opcode == TID_OP(READ_REQ)) {
2026 struct ib_reth *reth;
2034 reth = &ohdr->u.tid_rdma.r_req.reth;
2036 * The requester always restarts from the start of the original
2039 offset = delta_psn(psn, e->psn) * qp->pmtu;
2040 len = be32_to_cpu(reth->length);
2041 if (psn != e->psn || len != req->total_len)
2044 release_rdma_sge_mr(e);
2046 rkey = be32_to_cpu(reth->rkey);
2047 vaddr = get_ib_reth_vaddr(reth);
2050 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
2051 IB_ACCESS_REMOTE_READ);
2056 * If all the response packets for the current request have
2057 * been sent out and this request is complete (old_request
2058 * == false) and the TID flow may be unusable (the
2059 * req->clear_tail is advanced). However, when an earlier
2060 * request is received, this request will not be complete any
2061 * more (qp->s_tail_ack_queue is moved back, see below).
2062 * Consequently, we need to update the TID flow info everytime
2063 * a duplicate request is received.
2065 bth0 = be32_to_cpu(ohdr->bth[0]);
2066 if (tid_rdma_rcv_read_request(qp, e, packet, ohdr, bth0, psn,
2071 * True if the request is already scheduled (between
2072 * qp->s_tail_ack_queue and qp->r_head_ack_queue);
2077 struct flow_state *fstate;
2078 bool schedule = false;
2081 if (req->state == TID_REQUEST_RESEND) {
2082 req->state = TID_REQUEST_RESEND_ACTIVE;
2083 } else if (req->state == TID_REQUEST_INIT_RESEND) {
2084 req->state = TID_REQUEST_INIT;
2089 * True if the request is already scheduled (between
2090 * qp->s_tail_ack_queue and qp->r_head_ack_queue).
2091 * Also, don't change requests, which are at the SYNC
2092 * point and haven't generated any responses yet.
2093 * There is nothing to retransmit for them yet.
2095 if (old_req || req->state == TID_REQUEST_INIT ||
2096 (req->state == TID_REQUEST_SYNC && !req->cur_seg)) {
2097 for (i = prev + 1; ; i++) {
2098 if (i > rvt_size_atomic(&dev->rdi))
2100 if (i == qp->r_head_ack_queue)
2102 e = &qp->s_ack_queue[i];
2103 req = ack_to_tid_req(e);
2104 if (e->opcode == TID_OP(WRITE_REQ) &&
2105 req->state == TID_REQUEST_INIT)
2106 req->state = TID_REQUEST_INIT_RESEND;
2109 * If the state of the request has been changed,
2110 * the first leg needs to get scheduled in order to
2111 * pick up the change. Otherwise, normal response
2112 * processing should take care of it.
2119 * If there is no more allocated segment, just schedule the qp
2120 * without changing any state.
2122 if (req->clear_tail == req->setup_head)
2125 * If this request has sent responses for segments, which have
2126 * not received data yet (flow_idx != clear_tail), the flow_idx
2127 * pointer needs to be adjusted so the same responses can be
2130 if (CIRC_CNT(req->flow_idx, req->clear_tail, MAX_FLOWS)) {
2131 fstate = &req->flows[req->clear_tail].flow_state;
2132 qpriv->pending_tid_w_segs -=
2133 CIRC_CNT(req->flow_idx, req->clear_tail,
2136 CIRC_ADD(req->clear_tail,
2137 delta_psn(psn, fstate->resp_ib_psn),
2139 qpriv->pending_tid_w_segs +=
2140 delta_psn(psn, fstate->resp_ib_psn);
2142 * When flow_idx == setup_head, we've gotten a duplicate
2143 * request for a segment, which has not been allocated
2144 * yet. In that case, don't adjust this request.
2145 * However, we still want to go through the loop below
2146 * to adjust all subsequent requests.
2148 if (CIRC_CNT(req->setup_head, req->flow_idx,
2150 req->cur_seg = delta_psn(psn, e->psn);
2151 req->state = TID_REQUEST_RESEND_ACTIVE;
2155 for (i = prev + 1; ; i++) {
2157 * Look at everything up to and including
2160 if (i > rvt_size_atomic(&dev->rdi))
2162 if (i == qp->r_head_ack_queue)
2164 e = &qp->s_ack_queue[i];
2165 req = ack_to_tid_req(e);
2166 trace_hfi1_tid_req_rcv_err(qp, 0, e->opcode, e->psn,
2168 if (e->opcode != TID_OP(WRITE_REQ) ||
2169 req->cur_seg == req->comp_seg ||
2170 req->state == TID_REQUEST_INIT ||
2171 req->state == TID_REQUEST_INIT_RESEND) {
2172 if (req->state == TID_REQUEST_INIT)
2173 req->state = TID_REQUEST_INIT_RESEND;
2176 qpriv->pending_tid_w_segs -=
2177 CIRC_CNT(req->flow_idx,
2180 req->flow_idx = req->clear_tail;
2181 req->state = TID_REQUEST_RESEND;
2182 req->cur_seg = req->comp_seg;
2184 qpriv->s_flags &= ~HFI1_R_TID_WAIT_INTERLCK;
2186 /* Re-process old requests.*/
2187 if (qp->s_acked_ack_queue == qp->s_tail_ack_queue)
2188 qp->s_acked_ack_queue = prev;
2189 qp->s_tail_ack_queue = prev;
2191 * Since the qp->s_tail_ack_queue is modified, the
2192 * qp->s_ack_state must be changed to re-initialize
2193 * qp->s_ack_rdma_sge; Otherwise, we will end up in
2194 * wrong memory region.
2196 qp->s_ack_state = OP(ACKNOWLEDGE);
2199 * It's possible to receive a retry psn that is earlier than an RNRNAK
2200 * psn. In this case, the rnrnak state should be cleared.
2202 if (qpriv->rnr_nak_state) {
2203 qp->s_nak_state = 0;
2204 qpriv->rnr_nak_state = TID_RNR_NAK_INIT;
2205 qp->r_psn = e->lpsn + 1;
2206 hfi1_tid_write_alloc_resources(qp, true);
2209 qp->r_state = e->opcode;
2210 qp->r_nak_state = 0;
2211 qp->s_flags |= RVT_S_RESP_PENDING;
2212 hfi1_schedule_send(qp);
2214 spin_unlock_irqrestore(&qp->s_lock, flags);
2219 void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet)
2221 /* HANDLER FOR TID RDMA READ REQUEST packet (Responder side)*/
2224 * 1. Verify TID RDMA READ REQ as per IB_OPCODE_RC_RDMA_READ
2225 * (see hfi1_rc_rcv())
2226 * 2. Put TID RDMA READ REQ into the response queueu (s_ack_queue)
2227 * - Setup struct tid_rdma_req with request info
2228 * - Initialize struct tid_rdma_flow info;
2229 * - Copy TID entries;
2230 * 3. Set the qp->s_ack_state.
2231 * 4. Set RVT_S_RESP_PENDING in s_flags.
2232 * 5. Kick the send engine (hfi1_schedule_send())
2234 struct hfi1_ctxtdata *rcd = packet->rcd;
2235 struct rvt_qp *qp = packet->qp;
2236 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
2237 struct ib_other_headers *ohdr = packet->ohdr;
2238 struct rvt_ack_entry *e;
2239 unsigned long flags;
2240 struct ib_reth *reth;
2241 struct hfi1_qp_priv *qpriv = qp->priv;
2242 u32 bth0, psn, len, rkey;
2247 u8 nack_state = IB_NAK_INVALID_REQUEST;
2249 bth0 = be32_to_cpu(ohdr->bth[0]);
2250 if (hfi1_ruc_check_hdr(ibp, packet))
2253 fecn = process_ecn(qp, packet);
2254 psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
2255 trace_hfi1_rsp_rcv_tid_read_req(qp, psn);
2257 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
2260 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
2263 reth = &ohdr->u.tid_rdma.r_req.reth;
2264 vaddr = be64_to_cpu(reth->vaddr);
2265 len = be32_to_cpu(reth->length);
2266 /* The length needs to be in multiples of PAGE_SIZE */
2267 if (!len || len & ~PAGE_MASK || len > qpriv->tid_rdma.local.max_len)
2270 diff = delta_psn(psn, qp->r_psn);
2271 if (unlikely(diff)) {
2272 tid_rdma_rcv_err(packet, ohdr, qp, psn, diff, fecn);
2276 /* We've verified the request, insert it into the ack queue. */
2277 next = qp->r_head_ack_queue + 1;
2278 if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
2280 spin_lock_irqsave(&qp->s_lock, flags);
2281 if (unlikely(next == qp->s_tail_ack_queue)) {
2282 if (!qp->s_ack_queue[next].sent) {
2283 nack_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2284 goto nack_inv_unlock;
2286 update_ack_queue(qp, next);
2288 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2289 release_rdma_sge_mr(e);
2291 rkey = be32_to_cpu(reth->rkey);
2294 if (unlikely(!rvt_rkey_ok(qp, &e->rdma_sge, qp->r_len, vaddr,
2295 rkey, IB_ACCESS_REMOTE_READ)))
2298 /* Accept the request parameters */
2299 if (tid_rdma_rcv_read_request(qp, e, packet, ohdr, bth0, psn, vaddr,
2301 goto nack_inv_unlock;
2303 qp->r_state = e->opcode;
2304 qp->r_nak_state = 0;
2306 * We need to increment the MSN here instead of when we
2307 * finish sending the result since a duplicate request would
2308 * increment it more than once.
2311 qp->r_psn += e->lpsn - e->psn + 1;
2313 qp->r_head_ack_queue = next;
2316 * For all requests other than TID WRITE which are added to the ack
2317 * queue, qpriv->r_tid_alloc follows qp->r_head_ack_queue. It is ok to
2318 * do this because of interlocks between these and TID WRITE
2319 * requests. The same change has also been made in hfi1_rc_rcv().
2321 qpriv->r_tid_alloc = qp->r_head_ack_queue;
2323 /* Schedule the send tasklet. */
2324 qp->s_flags |= RVT_S_RESP_PENDING;
2326 qp->s_flags |= RVT_S_ECN;
2327 hfi1_schedule_send(qp);
2329 spin_unlock_irqrestore(&qp->s_lock, flags);
2333 spin_unlock_irqrestore(&qp->s_lock, flags);
2335 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2336 qp->r_nak_state = nack_state;
2337 qp->r_ack_psn = qp->r_psn;
2338 /* Queue NAK for later */
2339 rc_defered_ack(rcd, qp);
2342 spin_unlock_irqrestore(&qp->s_lock, flags);
2343 rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2344 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2345 qp->r_ack_psn = qp->r_psn;
2348 u32 hfi1_build_tid_rdma_read_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,
2349 struct ib_other_headers *ohdr, u32 *bth0,
2350 u32 *bth1, u32 *bth2, u32 *len, bool *last)
2352 struct hfi1_ack_priv *epriv = e->priv;
2353 struct tid_rdma_request *req = &epriv->tid_req;
2354 struct hfi1_qp_priv *qpriv = qp->priv;
2355 struct tid_rdma_flow *flow = &req->flows[req->clear_tail];
2356 u32 tidentry = flow->tid_entry[flow->tid_idx];
2357 u32 tidlen = EXP_TID_GET(tidentry, LEN) << PAGE_SHIFT;
2358 struct tid_rdma_read_resp *resp = &ohdr->u.tid_rdma.r_rsp;
2359 u32 next_offset, om = KDETH_OM_LARGE;
2362 struct tid_rdma_params *remote;
2364 *len = min_t(u32, qp->pmtu, tidlen - flow->tid_offset);
2366 next_offset = flow->tid_offset + *len;
2367 last_pkt = (flow->sent >= flow->length);
2369 trace_hfi1_tid_entry_build_read_resp(qp, flow->tid_idx, tidentry);
2370 trace_hfi1_tid_flow_build_read_resp(qp, req->clear_tail, flow);
2373 remote = rcu_dereference(qpriv->tid_rdma.remote);
2378 KDETH_RESET(resp->kdeth0, KVER, 0x1);
2379 KDETH_SET(resp->kdeth0, SH, !last_pkt);
2380 KDETH_SET(resp->kdeth0, INTR, !!(!last_pkt && remote->urg));
2381 KDETH_SET(resp->kdeth0, TIDCTRL, EXP_TID_GET(tidentry, CTRL));
2382 KDETH_SET(resp->kdeth0, TID, EXP_TID_GET(tidentry, IDX));
2383 KDETH_SET(resp->kdeth0, OM, om == KDETH_OM_LARGE);
2384 KDETH_SET(resp->kdeth0, OFFSET, flow->tid_offset / om);
2385 KDETH_RESET(resp->kdeth1, JKEY, remote->jkey);
2386 resp->verbs_qp = cpu_to_be32(qp->remote_qpn);
2389 resp->aeth = rvt_compute_aeth(qp);
2390 resp->verbs_psn = cpu_to_be32(mask_psn(flow->flow_state.ib_spsn +
2393 *bth0 = TID_OP(READ_RESP) << 24;
2394 *bth1 = flow->tid_qpn;
2395 *bth2 = mask_psn(((flow->flow_state.spsn + flow->pkt++) &
2396 HFI1_KDETH_BTH_SEQ_MASK) |
2397 (flow->flow_state.generation <<
2398 HFI1_KDETH_BTH_SEQ_SHIFT));
2401 /* Advance to next flow */
2402 req->clear_tail = (req->clear_tail + 1) &
2405 if (next_offset >= tidlen) {
2406 flow->tid_offset = 0;
2409 flow->tid_offset = next_offset;
2412 hdwords = sizeof(ohdr->u.tid_rdma.r_rsp) / sizeof(u32);
2418 static inline struct tid_rdma_request *
2419 find_tid_request(struct rvt_qp *qp, u32 psn, enum ib_wr_opcode opcode)
2420 __must_hold(&qp->s_lock)
2422 struct rvt_swqe *wqe;
2423 struct tid_rdma_request *req = NULL;
2426 end = qp->s_cur + 1;
2427 if (end == qp->s_size)
2429 for (i = qp->s_acked; i != end;) {
2430 wqe = rvt_get_swqe_ptr(qp, i);
2431 if (cmp_psn(psn, wqe->psn) >= 0 &&
2432 cmp_psn(psn, wqe->lpsn) <= 0) {
2433 if (wqe->wr.opcode == opcode)
2434 req = wqe_to_tid_req(wqe);
2437 if (++i == qp->s_size)
2444 void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet)
2446 /* HANDLER FOR TID RDMA READ RESPONSE packet (Requestor side */
2449 * 1. Find matching SWQE
2450 * 2. Check that the entire segment has been read.
2451 * 3. Remove HFI1_S_WAIT_TID_RESP from s_flags.
2452 * 4. Free the TID flow resources.
2453 * 5. Kick the send engine (hfi1_schedule_send())
2455 struct ib_other_headers *ohdr = packet->ohdr;
2456 struct rvt_qp *qp = packet->qp;
2457 struct hfi1_qp_priv *priv = qp->priv;
2458 struct hfi1_ctxtdata *rcd = packet->rcd;
2459 struct tid_rdma_request *req;
2460 struct tid_rdma_flow *flow;
2463 unsigned long flags;
2466 trace_hfi1_sender_rcv_tid_read_resp(qp);
2467 fecn = process_ecn(qp, packet);
2468 kpsn = mask_psn(be32_to_cpu(ohdr->bth[2]));
2469 aeth = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.aeth);
2470 opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff;
2472 spin_lock_irqsave(&qp->s_lock, flags);
2473 ipsn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn));
2474 req = find_tid_request(qp, ipsn, IB_WR_TID_RDMA_READ);
2478 flow = &req->flows[req->clear_tail];
2479 /* When header suppression is disabled */
2480 if (cmp_psn(ipsn, flow->flow_state.ib_lpsn)) {
2481 update_r_next_psn_fecn(packet, priv, rcd, flow, fecn);
2483 if (cmp_psn(kpsn, flow->flow_state.r_next_psn))
2485 flow->flow_state.r_next_psn = mask_psn(kpsn + 1);
2487 * Copy the payload to destination buffer if this packet is
2488 * delivered as an eager packet due to RSM rule and FECN.
2489 * The RSM rule selects FECN bit in BTH and SH bit in
2490 * KDETH header and therefore will not match the last
2491 * packet of each segment that has SH bit cleared.
2493 if (fecn && packet->etype == RHF_RCV_TYPE_EAGER) {
2494 struct rvt_sge_state ss;
2496 u32 tlen = packet->tlen;
2497 u16 hdrsize = packet->hlen;
2498 u8 pad = packet->pad;
2499 u8 extra_bytes = pad + packet->extra_byte +
2501 u32 pmtu = qp->pmtu;
2503 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
2505 len = restart_sge(&ss, req->e.swqe, ipsn, pmtu);
2506 if (unlikely(len < pmtu))
2508 rvt_copy_sge(qp, &ss, packet->payload, pmtu, false,
2510 /* Raise the sw sequence check flag for next packet */
2511 priv->s_flags |= HFI1_R_TID_SW_PSN;
2516 flow->flow_state.r_next_psn = mask_psn(kpsn + 1);
2518 priv->pending_tid_r_segs--;
2519 qp->s_num_rd_atomic--;
2520 if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
2521 !qp->s_num_rd_atomic) {
2522 qp->s_flags &= ~(RVT_S_WAIT_FENCE |
2524 hfi1_schedule_send(qp);
2526 if (qp->s_flags & RVT_S_WAIT_RDMAR) {
2527 qp->s_flags &= ~(RVT_S_WAIT_RDMAR | RVT_S_WAIT_ACK);
2528 hfi1_schedule_send(qp);
2531 trace_hfi1_ack(qp, ipsn);
2532 trace_hfi1_tid_req_rcv_read_resp(qp, 0, req->e.swqe->wr.opcode,
2533 req->e.swqe->psn, req->e.swqe->lpsn,
2535 trace_hfi1_tid_flow_rcv_read_resp(qp, req->clear_tail, flow);
2537 /* Release the tid resources */
2538 hfi1_kern_exp_rcv_clear(req);
2540 if (!do_rc_ack(qp, aeth, ipsn, opcode, 0, rcd))
2543 /* If not done yet, build next read request */
2544 if (++req->comp_seg >= req->total_segs) {
2546 req->state = TID_REQUEST_COMPLETE;
2550 * Clear the hw flow under two conditions:
2551 * 1. This request is a sync point and it is complete;
2552 * 2. Current request is completed and there are no more requests.
2554 if ((req->state == TID_REQUEST_SYNC &&
2555 req->comp_seg == req->cur_seg) ||
2556 priv->tid_r_comp == priv->tid_r_reqs) {
2557 hfi1_kern_clear_hw_flow(priv->rcd, qp);
2558 priv->s_flags &= ~HFI1_R_TID_SW_PSN;
2559 if (req->state == TID_REQUEST_SYNC)
2560 req->state = TID_REQUEST_ACTIVE;
2563 hfi1_schedule_send(qp);
2568 * The test indicates that the send engine has finished its cleanup
2569 * after sending the request and it's now safe to put the QP into error
2570 * state. However, if the wqe queue is empty (qp->s_acked == qp->s_tail
2571 * == qp->s_head), it would be unsafe to complete the wqe pointed by
2572 * qp->s_acked here. Putting the qp into error state will safely flush
2573 * all remaining requests.
2575 if (qp->s_last == qp->s_acked)
2576 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
2579 spin_unlock_irqrestore(&qp->s_lock, flags);
2582 void hfi1_kern_read_tid_flow_free(struct rvt_qp *qp)
2583 __must_hold(&qp->s_lock)
2585 u32 n = qp->s_acked;
2586 struct rvt_swqe *wqe;
2587 struct tid_rdma_request *req;
2588 struct hfi1_qp_priv *priv = qp->priv;
2590 lockdep_assert_held(&qp->s_lock);
2591 /* Free any TID entries */
2592 while (n != qp->s_tail) {
2593 wqe = rvt_get_swqe_ptr(qp, n);
2594 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
2595 req = wqe_to_tid_req(wqe);
2596 hfi1_kern_exp_rcv_clear_all(req);
2599 if (++n == qp->s_size)
2603 hfi1_kern_clear_hw_flow(priv->rcd, qp);
2606 static bool tid_rdma_tid_err(struct hfi1_ctxtdata *rcd,
2607 struct hfi1_packet *packet, u8 rcv_type,
2610 struct rvt_qp *qp = packet->qp;
2611 struct hfi1_qp_priv *qpriv = qp->priv;
2613 struct ib_other_headers *ohdr = packet->ohdr;
2614 struct rvt_ack_entry *e;
2615 struct tid_rdma_request *req;
2616 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
2619 if (rcv_type >= RHF_RCV_TYPE_IB)
2622 spin_lock(&qp->s_lock);
2625 * We've ran out of space in the eager buffer.
2626 * Eagerly received KDETH packets which require space in the
2627 * Eager buffer (packet that have payload) are TID RDMA WRITE
2628 * response packets. In this case, we have to re-transmit the
2629 * TID RDMA WRITE request.
2631 if (rcv_type == RHF_RCV_TYPE_EAGER) {
2632 hfi1_restart_rc(qp, qp->s_last_psn + 1, 1);
2633 hfi1_schedule_send(qp);
2638 * For TID READ response, error out QP after freeing the tid
2641 if (opcode == TID_OP(READ_RESP)) {
2642 ipsn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn));
2643 if (cmp_psn(ipsn, qp->s_last_psn) > 0 &&
2644 cmp_psn(ipsn, qp->s_psn) < 0) {
2645 hfi1_kern_read_tid_flow_free(qp);
2646 spin_unlock(&qp->s_lock);
2647 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2654 * Error out the qp for TID RDMA WRITE
2656 hfi1_kern_clear_hw_flow(qpriv->rcd, qp);
2657 for (i = 0; i < rvt_max_atomic(rdi); i++) {
2658 e = &qp->s_ack_queue[i];
2659 if (e->opcode == TID_OP(WRITE_REQ)) {
2660 req = ack_to_tid_req(e);
2661 hfi1_kern_exp_rcv_clear_all(req);
2664 spin_unlock(&qp->s_lock);
2665 rvt_rc_error(qp, IB_WC_LOC_LEN_ERR);
2669 spin_unlock(&qp->s_lock);
2674 static void restart_tid_rdma_read_req(struct hfi1_ctxtdata *rcd,
2675 struct rvt_qp *qp, struct rvt_swqe *wqe)
2677 struct tid_rdma_request *req;
2678 struct tid_rdma_flow *flow;
2680 /* Start from the right segment */
2681 qp->r_flags |= RVT_R_RDMAR_SEQ;
2682 req = wqe_to_tid_req(wqe);
2683 flow = &req->flows[req->clear_tail];
2684 hfi1_restart_rc(qp, flow->flow_state.ib_spsn, 0);
2685 if (list_empty(&qp->rspwait)) {
2686 qp->r_flags |= RVT_R_RSP_SEND;
2688 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2693 * Handle the KDETH eflags for TID RDMA READ response.
2695 * Return true if the last packet for a segment has been received and it is
2696 * time to process the response normally; otherwise, return true.
2698 * The caller must hold the packet->qp->r_lock and the rcu_read_lock.
2700 static bool handle_read_kdeth_eflags(struct hfi1_ctxtdata *rcd,
2701 struct hfi1_packet *packet, u8 rcv_type,
2702 u8 rte, u32 psn, u32 ibpsn)
2703 __must_hold(&packet->qp->r_lock) __must_hold(RCU)
2705 struct hfi1_pportdata *ppd = rcd->ppd;
2706 struct hfi1_devdata *dd = ppd->dd;
2707 struct hfi1_ibport *ibp;
2708 struct rvt_swqe *wqe;
2709 struct tid_rdma_request *req;
2710 struct tid_rdma_flow *flow;
2712 struct rvt_qp *qp = packet->qp;
2713 struct hfi1_qp_priv *priv = qp->priv;
2718 lockdep_assert_held(&qp->r_lock);
2719 /* If the psn is out of valid range, drop the packet */
2720 if (cmp_psn(ibpsn, qp->s_last_psn) < 0 ||
2721 cmp_psn(ibpsn, qp->s_psn) > 0)
2724 spin_lock(&qp->s_lock);
2726 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
2727 * requests and implicitly NAK RDMA read and atomic requests issued
2728 * before the NAK'ed request.
2730 ack_psn = ibpsn - 1;
2731 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2732 ibp = to_iport(qp->ibqp.device, qp->port_num);
2734 /* Complete WQEs that the PSN finishes. */
2735 while ((int)delta_psn(ack_psn, wqe->lpsn) >= 0) {
2737 * If this request is a RDMA read or atomic, and the NACK is
2738 * for a later operation, this NACK NAKs the RDMA read or
2741 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
2742 wqe->wr.opcode == IB_WR_TID_RDMA_READ ||
2743 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2744 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2745 /* Retry this request. */
2746 if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
2747 qp->r_flags |= RVT_R_RDMAR_SEQ;
2748 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
2749 restart_tid_rdma_read_req(rcd, qp,
2752 hfi1_restart_rc(qp, qp->s_last_psn + 1,
2754 if (list_empty(&qp->rspwait)) {
2755 qp->r_flags |= RVT_R_RSP_SEND;
2757 list_add_tail(/* wait */
2759 &rcd->qp_wait_list);
2764 * No need to process the NAK since we are
2765 * restarting an earlier request.
2770 wqe = do_rc_completion(qp, wqe, ibp);
2771 if (qp->s_acked == qp->s_tail)
2775 /* Handle the eflags for the request */
2776 if (wqe->wr.opcode != IB_WR_TID_RDMA_READ)
2779 req = wqe_to_tid_req(wqe);
2781 case RHF_RCV_TYPE_EXPECTED:
2783 case RHF_RTE_EXPECTED_FLOW_SEQ_ERR:
2785 * On the first occurrence of a Flow Sequence error,
2786 * the flag TID_FLOW_SW_PSN is set.
2788 * After that, the flow is *not* reprogrammed and the
2789 * protocol falls back to SW PSN checking. This is done
2790 * to prevent continuous Flow Sequence errors for any
2791 * packets that could be still in the fabric.
2793 flow = find_flow(req, psn, NULL);
2796 * We can't find the IB PSN matching the
2797 * received KDETH PSN. The only thing we can
2798 * do at this point is report the error to
2801 hfi1_kern_read_tid_flow_free(qp);
2802 spin_unlock(&qp->s_lock);
2803 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2806 if (priv->s_flags & HFI1_R_TID_SW_PSN) {
2808 flow->flow_state.r_next_psn);
2810 if (!(qp->r_flags & RVT_R_RDMAR_SEQ))
2811 restart_tid_rdma_read_req(rcd,
2815 /* Drop the packet.*/
2817 } else if (diff < 0) {
2819 * If a response packet for a restarted
2820 * request has come back, reset the
2823 if (qp->r_flags & RVT_R_RDMAR_SEQ)
2827 /* Drop the packet.*/
2832 * If SW PSN verification is successful and
2833 * this is the last packet in the segment, tell
2834 * the caller to process it as a normal packet.
2836 fpsn = full_flow_psn(flow,
2837 flow->flow_state.lpsn);
2838 if (cmp_psn(fpsn, psn) == 0) {
2840 if (qp->r_flags & RVT_R_RDMAR_SEQ)
2844 flow->flow_state.r_next_psn =
2849 last_psn = read_r_next_psn(dd, rcd->ctxt,
2851 flow->flow_state.r_next_psn = last_psn;
2852 priv->s_flags |= HFI1_R_TID_SW_PSN;
2854 * If no request has been restarted yet,
2855 * restart the current one.
2857 if (!(qp->r_flags & RVT_R_RDMAR_SEQ))
2858 restart_tid_rdma_read_req(rcd, qp,
2864 case RHF_RTE_EXPECTED_FLOW_GEN_ERR:
2866 * Since the TID flow is able to ride through
2867 * generation mismatch, drop this stale packet.
2876 case RHF_RCV_TYPE_ERROR:
2878 case RHF_RTE_ERROR_OP_CODE_ERR:
2879 case RHF_RTE_ERROR_KHDR_MIN_LEN_ERR:
2880 case RHF_RTE_ERROR_KHDR_HCRC_ERR:
2881 case RHF_RTE_ERROR_KHDR_KVER_ERR:
2882 case RHF_RTE_ERROR_CONTEXT_ERR:
2883 case RHF_RTE_ERROR_KHDR_TID_ERR:
2891 spin_unlock(&qp->s_lock);
2895 bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd,
2896 struct hfi1_pportdata *ppd,
2897 struct hfi1_packet *packet)
2899 struct hfi1_ibport *ibp = &ppd->ibport_data;
2900 struct hfi1_devdata *dd = ppd->dd;
2901 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
2902 u8 rcv_type = rhf_rcv_type(packet->rhf);
2903 u8 rte = rhf_rcv_type_err(packet->rhf);
2904 struct ib_header *hdr = packet->hdr;
2905 struct ib_other_headers *ohdr = NULL;
2906 int lnh = be16_to_cpu(hdr->lrh[0]) & 3;
2907 u16 lid = be16_to_cpu(hdr->lrh[1]);
2909 u32 qp_num, psn, ibpsn;
2911 struct hfi1_qp_priv *qpriv;
2912 unsigned long flags;
2914 struct rvt_ack_entry *e;
2915 struct tid_rdma_request *req;
2916 struct tid_rdma_flow *flow;
2919 trace_hfi1_msg_handle_kdeth_eflags(NULL, "Kdeth error: rhf ",
2921 if (packet->rhf & RHF_ICRC_ERR)
2924 packet->ohdr = &hdr->u.oth;
2925 ohdr = packet->ohdr;
2926 trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf)));
2928 /* Get the destination QP number. */
2929 qp_num = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_qp) &
2931 if (lid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
2934 psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
2935 opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff;
2938 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
2944 /* Check for valid receive state. */
2945 spin_lock_irqsave(&qp->r_lock, flags);
2946 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
2947 ibp->rvp.n_pkt_drops++;
2951 if (packet->rhf & RHF_TID_ERR) {
2952 /* For TIDERR and RC QPs preemptively schedule a NAK */
2953 u32 tlen = rhf_pkt_len(packet->rhf); /* in bytes */
2955 /* Sanity check packet */
2960 * Check for GRH. We should never get packets with GRH in this
2963 if (lnh == HFI1_LRH_GRH)
2966 if (tid_rdma_tid_err(rcd, packet, rcv_type, opcode))
2970 /* handle TID RDMA READ */
2971 if (opcode == TID_OP(READ_RESP)) {
2972 ibpsn = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn);
2973 ibpsn = mask_psn(ibpsn);
2974 ret = handle_read_kdeth_eflags(rcd, packet, rcv_type, rte, psn,
2980 * qp->s_tail_ack_queue points to the rvt_ack_entry currently being
2981 * processed. These a completed sequentially so we can be sure that
2982 * the pointer will not change until the entire request has completed.
2984 spin_lock(&qp->s_lock);
2986 e = &qp->s_ack_queue[qpriv->r_tid_tail];
2987 req = ack_to_tid_req(e);
2988 flow = &req->flows[req->clear_tail];
2989 trace_hfi1_eflags_err_write(qp, rcv_type, rte, psn);
2990 trace_hfi1_rsp_handle_kdeth_eflags(qp, psn);
2991 trace_hfi1_tid_write_rsp_handle_kdeth_eflags(qp);
2992 trace_hfi1_tid_req_handle_kdeth_eflags(qp, 0, e->opcode, e->psn,
2994 trace_hfi1_tid_flow_handle_kdeth_eflags(qp, req->clear_tail, flow);
2997 case RHF_RCV_TYPE_EXPECTED:
2999 case RHF_RTE_EXPECTED_FLOW_SEQ_ERR:
3000 if (!(qpriv->s_flags & HFI1_R_TID_SW_PSN)) {
3001 qpriv->s_flags |= HFI1_R_TID_SW_PSN;
3002 flow->flow_state.r_next_psn =
3003 read_r_next_psn(dd, rcd->ctxt,
3005 qpriv->r_next_psn_kdeth =
3006 flow->flow_state.r_next_psn;
3010 * If the received PSN does not match the next
3011 * expected PSN, NAK the packet.
3012 * However, only do that if we know that the a
3013 * NAK has already been sent. Otherwise, this
3014 * mismatch could be due to packets that were
3015 * already in flight.
3018 flow->flow_state.r_next_psn);
3024 qpriv->s_nak_state = 0;
3026 * If SW PSN verification is successful and this
3027 * is the last packet in the segment, tell the
3028 * caller to process it as a normal packet.
3030 if (psn == full_flow_psn(flow,
3031 flow->flow_state.lpsn))
3033 flow->flow_state.r_next_psn =
3035 qpriv->r_next_psn_kdeth =
3036 flow->flow_state.r_next_psn;
3040 case RHF_RTE_EXPECTED_FLOW_GEN_ERR:
3048 case RHF_RCV_TYPE_ERROR:
3050 case RHF_RTE_ERROR_OP_CODE_ERR:
3051 case RHF_RTE_ERROR_KHDR_MIN_LEN_ERR:
3052 case RHF_RTE_ERROR_KHDR_HCRC_ERR:
3053 case RHF_RTE_ERROR_KHDR_KVER_ERR:
3054 case RHF_RTE_ERROR_CONTEXT_ERR:
3055 case RHF_RTE_ERROR_KHDR_TID_ERR:
3064 spin_unlock(&qp->s_lock);
3066 spin_unlock_irqrestore(&qp->r_lock, flags);
3072 ibp->rvp.n_rc_seqnak++;
3073 if (!qpriv->s_nak_state) {
3074 qpriv->s_nak_state = IB_NAK_PSN_ERROR;
3075 /* We are NAK'ing the next expected PSN */
3076 qpriv->s_nak_psn = mask_psn(flow->flow_state.r_next_psn);
3077 qpriv->s_flags |= RVT_S_ACK_PENDING;
3078 if (qpriv->r_tid_ack == HFI1_QP_WQE_INVALID)
3079 qpriv->r_tid_ack = qpriv->r_tid_tail;
3080 hfi1_schedule_tid_send(qp);
3086 * "Rewind" the TID request information.
3087 * This means that we reset the state back to ACTIVE,
3088 * find the proper flow, set the flow index to that flow,
3089 * and reset the flow information.
3091 void hfi1_tid_rdma_restart_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
3094 struct tid_rdma_request *req = wqe_to_tid_req(wqe);
3095 struct tid_rdma_flow *flow;
3096 struct hfi1_qp_priv *qpriv = qp->priv;
3097 int diff, delta_pkts;
3101 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
3102 *bth2 = mask_psn(qp->s_psn);
3103 flow = find_flow_ib(req, *bth2, &fidx);
3105 trace_hfi1_msg_tid_restart_req(/* msg */
3106 qp, "!!!!!! Could not find flow to restart: bth2 ",
3108 trace_hfi1_tid_req_restart_req(qp, 0, wqe->wr.opcode,
3109 wqe->psn, wqe->lpsn,
3114 fidx = req->acked_tail;
3115 flow = &req->flows[fidx];
3116 *bth2 = mask_psn(req->r_ack_psn);
3119 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
3120 delta_pkts = delta_psn(*bth2, flow->flow_state.ib_spsn);
3122 delta_pkts = delta_psn(*bth2,
3124 flow->flow_state.spsn));
3126 trace_hfi1_tid_flow_restart_req(qp, fidx, flow);
3127 diff = delta_pkts + flow->resync_npkts;
3132 flow->tid_offset = 0;
3134 for (tididx = 0; tididx < flow->tidcnt; tididx++) {
3135 u32 tidentry = flow->tid_entry[tididx], tidlen,
3138 flow->tid_offset = 0;
3139 tidlen = EXP_TID_GET(tidentry, LEN) * PAGE_SIZE;
3140 tidnpkts = rvt_div_round_up_mtu(qp, tidlen);
3141 npkts = min_t(u32, diff, tidnpkts);
3143 flow->sent += (npkts == tidnpkts ? tidlen :
3145 flow->tid_offset += npkts * qp->pmtu;
3151 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE) {
3152 rvt_skip_sge(&qpriv->tid_ss, (req->cur_seg * req->seg_len) +
3155 * Packet PSN is based on flow_state.spsn + flow->pkt. However,
3156 * during a RESYNC, the generation is incremented and the
3157 * sequence is reset to 0. Since we've adjusted the npkts in the
3158 * flow and the SGE has been sufficiently advanced, we have to
3159 * adjust flow->pkt in order to calculate the correct PSN.
3161 flow->pkt -= flow->resync_npkts;
3164 if (flow->tid_offset ==
3165 EXP_TID_GET(flow->tid_entry[tididx], LEN) * PAGE_SIZE) {
3167 flow->tid_offset = 0;
3169 flow->tid_idx = tididx;
3170 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
3171 /* Move flow_idx to correct index */
3172 req->flow_idx = fidx;
3174 req->clear_tail = fidx;
3176 trace_hfi1_tid_flow_restart_req(qp, fidx, flow);
3177 trace_hfi1_tid_req_restart_req(qp, 0, wqe->wr.opcode, wqe->psn,
3179 req->state = TID_REQUEST_ACTIVE;
3180 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE) {
3181 /* Reset all the flows that we are going to resend */
3182 fidx = CIRC_NEXT(fidx, MAX_FLOWS);
3183 i = qpriv->s_tid_tail;
3185 for (; CIRC_CNT(req->setup_head, fidx, MAX_FLOWS);
3186 fidx = CIRC_NEXT(fidx, MAX_FLOWS)) {
3187 req->flows[fidx].sent = 0;
3188 req->flows[fidx].pkt = 0;
3189 req->flows[fidx].tid_idx = 0;
3190 req->flows[fidx].tid_offset = 0;
3191 req->flows[fidx].resync_npkts = 0;
3193 if (i == qpriv->s_tid_cur)
3196 i = (++i == qp->s_size ? 0 : i);
3197 wqe = rvt_get_swqe_ptr(qp, i);
3198 } while (wqe->wr.opcode != IB_WR_TID_RDMA_WRITE);
3199 req = wqe_to_tid_req(wqe);
3200 req->cur_seg = req->ack_seg;
3201 fidx = req->acked_tail;
3202 /* Pull req->clear_tail back */
3203 req->clear_tail = fidx;
3208 void hfi1_qp_kern_exp_rcv_clear_all(struct rvt_qp *qp)
3211 struct hfi1_qp_priv *qpriv = qp->priv;
3212 struct tid_flow_state *fs;
3214 if (qp->ibqp.qp_type != IB_QPT_RC || !HFI1_CAP_IS_KSET(TID_RDMA))
3218 * First, clear the flow to help prevent any delayed packets from
3221 fs = &qpriv->flow_state;
3222 if (fs->index != RXE_NUM_TID_FLOWS)
3223 hfi1_kern_clear_hw_flow(qpriv->rcd, qp);
3225 for (i = qp->s_acked; i != qp->s_head;) {
3226 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, i);
3228 if (++i == qp->s_size)
3230 /* Free only locally allocated TID entries */
3231 if (wqe->wr.opcode != IB_WR_TID_RDMA_READ)
3234 struct hfi1_swqe_priv *priv = wqe->priv;
3236 ret = hfi1_kern_exp_rcv_clear(&priv->tid_req);
3239 for (i = qp->s_acked_ack_queue; i != qp->r_head_ack_queue;) {
3240 struct rvt_ack_entry *e = &qp->s_ack_queue[i];
3242 if (++i == rvt_max_atomic(ib_to_rvt(qp->ibqp.device)))
3244 /* Free only locally allocated TID entries */
3245 if (e->opcode != TID_OP(WRITE_REQ))
3248 struct hfi1_ack_priv *priv = e->priv;
3250 ret = hfi1_kern_exp_rcv_clear(&priv->tid_req);
3255 bool hfi1_tid_rdma_wqe_interlock(struct rvt_qp *qp, struct rvt_swqe *wqe)
3257 struct rvt_swqe *prev;
3258 struct hfi1_qp_priv *priv = qp->priv;
3260 struct tid_rdma_request *req;
3262 s_prev = (qp->s_cur == 0 ? qp->s_size : qp->s_cur) - 1;
3263 prev = rvt_get_swqe_ptr(qp, s_prev);
3265 switch (wqe->wr.opcode) {
3267 case IB_WR_SEND_WITH_IMM:
3268 case IB_WR_SEND_WITH_INV:
3269 case IB_WR_ATOMIC_CMP_AND_SWP:
3270 case IB_WR_ATOMIC_FETCH_AND_ADD:
3271 case IB_WR_RDMA_WRITE:
3272 switch (prev->wr.opcode) {
3273 case IB_WR_TID_RDMA_WRITE:
3274 req = wqe_to_tid_req(prev);
3275 if (req->ack_seg != req->total_segs)
3281 case IB_WR_RDMA_READ:
3282 if (prev->wr.opcode != IB_WR_TID_RDMA_WRITE)
3285 case IB_WR_TID_RDMA_READ:
3286 switch (prev->wr.opcode) {
3287 case IB_WR_RDMA_READ:
3288 if (qp->s_acked != qp->s_cur)
3291 case IB_WR_TID_RDMA_WRITE:
3292 req = wqe_to_tid_req(prev);
3293 if (req->ack_seg != req->total_segs)
3304 priv->s_flags |= HFI1_S_TID_WAIT_INTERLCK;
3308 /* Does @sge meet the alignment requirements for tid rdma? */
3309 static inline bool hfi1_check_sge_align(struct rvt_qp *qp,
3310 struct rvt_sge *sge, int num_sge)
3314 for (i = 0; i < num_sge; i++, sge++) {
3315 trace_hfi1_sge_check_align(qp, i, sge);
3316 if ((u64)sge->vaddr & ~PAGE_MASK ||
3317 sge->sge_length & ~PAGE_MASK)
3323 void setup_tid_rdma_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe)
3325 struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv;
3326 struct hfi1_swqe_priv *priv = wqe->priv;
3327 struct tid_rdma_params *remote;
3328 enum ib_wr_opcode new_opcode;
3329 bool do_tid_rdma = false;
3330 struct hfi1_pportdata *ppd = qpriv->rcd->ppd;
3332 if ((rdma_ah_get_dlid(&qp->remote_ah_attr) & ~((1 << ppd->lmc) - 1)) ==
3335 if (qpriv->hdr_type != HFI1_PKT_TYPE_9B)
3339 remote = rcu_dereference(qpriv->tid_rdma.remote);
3341 * If TID RDMA is disabled by the negotiation, don't
3347 if (wqe->wr.opcode == IB_WR_RDMA_READ) {
3348 if (hfi1_check_sge_align(qp, &wqe->sg_list[0],
3350 new_opcode = IB_WR_TID_RDMA_READ;
3353 } else if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
3355 * TID RDMA is enabled for this RDMA WRITE request iff:
3356 * 1. The remote address is page-aligned,
3357 * 2. The length is larger than the minimum segment size,
3358 * 3. The length is page-multiple.
3360 if (!(wqe->rdma_wr.remote_addr & ~PAGE_MASK) &&
3361 !(wqe->length & ~PAGE_MASK)) {
3362 new_opcode = IB_WR_TID_RDMA_WRITE;
3368 if (hfi1_kern_exp_rcv_alloc_flows(&priv->tid_req, GFP_ATOMIC))
3370 wqe->wr.opcode = new_opcode;
3371 priv->tid_req.seg_len =
3372 min_t(u32, remote->max_len, wqe->length);
3373 priv->tid_req.total_segs =
3374 DIV_ROUND_UP(wqe->length, priv->tid_req.seg_len);
3375 /* Compute the last PSN of the request */
3376 wqe->lpsn = wqe->psn;
3377 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
3378 priv->tid_req.n_flows = remote->max_read;
3379 qpriv->tid_r_reqs++;
3380 wqe->lpsn += rvt_div_round_up_mtu(qp, wqe->length) - 1;
3382 wqe->lpsn += priv->tid_req.total_segs - 1;
3383 atomic_inc(&qpriv->n_requests);
3386 priv->tid_req.cur_seg = 0;
3387 priv->tid_req.comp_seg = 0;
3388 priv->tid_req.ack_seg = 0;
3389 priv->tid_req.state = TID_REQUEST_INACTIVE;
3392 * TID RDMA READ does not have ACKs so it does not
3393 * update the pointer. We have to reset it so TID RDMA
3394 * WRITE does not get confused.
3396 priv->tid_req.acked_tail = priv->tid_req.setup_head;
3397 trace_hfi1_tid_req_setup_tid_wqe(qp, 1, wqe->wr.opcode,
3398 wqe->psn, wqe->lpsn,
3405 /* TID RDMA WRITE functions */
3407 u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
3408 struct ib_other_headers *ohdr,
3409 u32 *bth1, u32 *bth2, u32 *len)
3411 struct hfi1_qp_priv *qpriv = qp->priv;
3412 struct tid_rdma_request *req = wqe_to_tid_req(wqe);
3413 struct tid_rdma_params *remote;
3416 remote = rcu_dereference(qpriv->tid_rdma.remote);
3418 * Set the number of flow to be used based on negotiated
3421 req->n_flows = remote->max_write;
3422 req->state = TID_REQUEST_ACTIVE;
3424 KDETH_RESET(ohdr->u.tid_rdma.w_req.kdeth0, KVER, 0x1);
3425 KDETH_RESET(ohdr->u.tid_rdma.w_req.kdeth1, JKEY, remote->jkey);
3426 ohdr->u.tid_rdma.w_req.reth.vaddr =
3427 cpu_to_be64(wqe->rdma_wr.remote_addr + (wqe->length - *len));
3428 ohdr->u.tid_rdma.w_req.reth.rkey =
3429 cpu_to_be32(wqe->rdma_wr.rkey);
3430 ohdr->u.tid_rdma.w_req.reth.length = cpu_to_be32(*len);
3431 ohdr->u.tid_rdma.w_req.verbs_qp = cpu_to_be32(qp->remote_qpn);
3432 *bth1 &= ~RVT_QPN_MASK;
3433 *bth1 |= remote->qp;
3434 qp->s_state = TID_OP(WRITE_REQ);
3435 qp->s_flags |= HFI1_S_WAIT_TID_RESP;
3436 *bth2 |= IB_BTH_REQ_ACK;
3440 return sizeof(ohdr->u.tid_rdma.w_req) / sizeof(u32);
3443 void hfi1_compute_tid_rdma_flow_wt(void)
3446 * Heuristic for computing the RNR timeout when waiting on the flow
3447 * queue. Rather than a computationaly expensive exact estimate of when
3448 * a flow will be available, we assume that if a QP is at position N in
3449 * the flow queue it has to wait approximately (N + 1) * (number of
3450 * segments between two sync points), assuming PMTU of 4K. The rationale
3451 * for this is that flows are released and recycled at each sync point.
3453 tid_rdma_flow_wt = MAX_TID_FLOW_PSN * enum_to_mtu(OPA_MTU_4096) /
3454 TID_RDMA_MAX_SEGMENT_SIZE;
3457 static u32 position_in_queue(struct hfi1_qp_priv *qpriv,
3458 struct tid_queue *queue)
3460 return qpriv->tid_enqueue - queue->dequeue;
3464 * @qp: points to rvt_qp context.
3465 * @to_seg: desired RNR timeout in segments.
3466 * Return: index of the next highest timeout in the ib_hfi1_rnr_table[]
3468 static u32 hfi1_compute_tid_rnr_timeout(struct rvt_qp *qp, u32 to_seg)
3470 struct hfi1_qp_priv *qpriv = qp->priv;
3475 bytes_per_us = active_egress_rate(qpriv->rcd->ppd) / 8;
3476 timeout = (to_seg * TID_RDMA_MAX_SEGMENT_SIZE) / bytes_per_us;
3478 * Find the next highest value in the RNR table to the required
3479 * timeout. This gives the responder some padding.
3481 for (i = 1; i <= IB_AETH_CREDIT_MASK; i++)
3482 if (rvt_rnr_tbl_to_usec(i) >= timeout)
3488 * Central place for resource allocation at TID write responder,
3489 * is called from write_req and write_data interrupt handlers as
3490 * well as the send thread when a queued QP is scheduled for
3491 * resource allocation.
3493 * Iterates over (a) segments of a request and then (b) queued requests
3494 * themselves to allocate resources for up to local->max_write
3495 * segments across multiple requests. Stop allocating when we
3496 * hit a sync point, resume allocating after data packets at
3497 * sync point have been received.
3499 * Resource allocation and sending of responses is decoupled. The
3500 * request/segment which are being allocated and sent are as follows.
3501 * Resources are allocated for:
3502 * [request: qpriv->r_tid_alloc, segment: req->alloc_seg]
3503 * The send thread sends:
3504 * [request: qp->s_tail_ack_queue, segment:req->cur_seg]
3506 static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx)
3508 struct tid_rdma_request *req;
3509 struct hfi1_qp_priv *qpriv = qp->priv;
3510 struct hfi1_ctxtdata *rcd = qpriv->rcd;
3511 struct tid_rdma_params *local = &qpriv->tid_rdma.local;
3512 struct rvt_ack_entry *e;
3517 lockdep_assert_held(&qp->s_lock);
3520 trace_hfi1_rsp_tid_write_alloc_res(qp, 0);
3521 trace_hfi1_tid_write_rsp_alloc_res(qp);
3523 * Don't allocate more segments if a RNR NAK has already been
3524 * scheduled to avoid messing up qp->r_psn: the RNR NAK will
3525 * be sent only when all allocated segments have been sent.
3526 * However, if more segments are allocated before that, TID RDMA
3527 * WRITE RESP packets will be sent out for these new segments
3528 * before the RNR NAK packet. When the requester receives the
3529 * RNR NAK packet, it will restart with qp->s_last_psn + 1,
3530 * which does not match qp->r_psn and will be dropped.
3531 * Consequently, the requester will exhaust its retries and
3532 * put the qp into error state.
3534 if (qpriv->rnr_nak_state == TID_RNR_NAK_SEND)
3537 /* No requests left to process */
3538 if (qpriv->r_tid_alloc == qpriv->r_tid_head) {
3539 /* If all data has been received, clear the flow */
3540 if (qpriv->flow_state.index < RXE_NUM_TID_FLOWS &&
3541 !qpriv->alloc_w_segs) {
3542 hfi1_kern_clear_hw_flow(rcd, qp);
3543 qpriv->s_flags &= ~HFI1_R_TID_SW_PSN;
3548 e = &qp->s_ack_queue[qpriv->r_tid_alloc];
3549 if (e->opcode != TID_OP(WRITE_REQ))
3551 req = ack_to_tid_req(e);
3552 trace_hfi1_tid_req_write_alloc_res(qp, 0, e->opcode, e->psn,
3554 /* Finished allocating for all segments of this request */
3555 if (req->alloc_seg >= req->total_segs)
3558 /* Can allocate only a maximum of local->max_write for a QP */
3559 if (qpriv->alloc_w_segs >= local->max_write)
3562 /* Don't allocate at a sync point with data packets pending */
3563 if (qpriv->sync_pt && qpriv->alloc_w_segs)
3566 /* All data received at the sync point, continue */
3567 if (qpriv->sync_pt && !qpriv->alloc_w_segs) {
3568 hfi1_kern_clear_hw_flow(rcd, qp);
3569 qpriv->sync_pt = false;
3570 qpriv->s_flags &= ~HFI1_R_TID_SW_PSN;
3573 /* Allocate flow if we don't have one */
3574 if (qpriv->flow_state.index >= RXE_NUM_TID_FLOWS) {
3575 ret = hfi1_kern_setup_hw_flow(qpriv->rcd, qp);
3577 to_seg = tid_rdma_flow_wt *
3578 position_in_queue(qpriv,
3584 npkts = rvt_div_round_up_mtu(qp, req->seg_len);
3587 * We are at a sync point if we run out of KDETH PSN space.
3588 * Last PSN of every generation is reserved for RESYNC.
3590 if (qpriv->flow_state.psn + npkts > MAX_TID_FLOW_PSN - 1) {
3591 qpriv->sync_pt = true;
3596 * If overtaking req->acked_tail, send an RNR NAK. Because the
3597 * QP is not queued in this case, and the issue can only be
3598 * caused due a delay in scheduling the second leg which we
3599 * cannot estimate, we use a rather arbitrary RNR timeout of
3600 * (MAX_FLOWS / 2) segments
3602 if (!CIRC_SPACE(req->setup_head, req->acked_tail,
3605 to_seg = MAX_FLOWS >> 1;
3606 qpriv->s_flags |= RVT_S_ACK_PENDING;
3607 hfi1_schedule_tid_send(qp);
3611 /* Try to allocate rcv array / TID entries */
3612 ret = hfi1_kern_exp_rcv_setup(req, &req->ss, &last);
3614 to_seg = position_in_queue(qpriv, &rcd->rarr_queue);
3618 qpriv->alloc_w_segs++;
3622 /* Begin processing the next request */
3623 if (++qpriv->r_tid_alloc >
3624 rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
3625 qpriv->r_tid_alloc = 0;
3629 * Schedule an RNR NAK to be sent if (a) flow or rcv array allocation
3630 * has failed (b) we are called from the rcv handler interrupt context
3631 * (c) an RNR NAK has not already been scheduled
3633 if (ret == -EAGAIN && intr_ctx && !qp->r_nak_state)
3639 lockdep_assert_held(&qp->r_lock);
3641 /* Set r_nak_state to prevent unrelated events from generating NAK's */
3642 qp->r_nak_state = hfi1_compute_tid_rnr_timeout(qp, to_seg) | IB_RNR_NAK;
3644 /* Pull back r_psn to the segment being RNR NAK'd */
3645 qp->r_psn = e->psn + req->alloc_seg;
3646 qp->r_ack_psn = qp->r_psn;
3648 * Pull back r_head_ack_queue to the ack entry following the request
3649 * being RNR NAK'd. This allows resources to be allocated to the request
3650 * if the queued QP is scheduled.
3652 qp->r_head_ack_queue = qpriv->r_tid_alloc + 1;
3653 if (qp->r_head_ack_queue > rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
3654 qp->r_head_ack_queue = 0;
3655 qpriv->r_tid_head = qp->r_head_ack_queue;
3657 * These send side fields are used in make_rc_ack(). They are set in
3658 * hfi1_send_rc_ack() but must be set here before dropping qp->s_lock
3661 qp->s_nak_state = qp->r_nak_state;
3662 qp->s_ack_psn = qp->r_ack_psn;
3664 * Clear the ACK PENDING flag to prevent unwanted ACK because we
3665 * have modified qp->s_ack_psn here.
3667 qp->s_flags &= ~(RVT_S_ACK_PENDING);
3669 trace_hfi1_rsp_tid_write_alloc_res(qp, qp->r_psn);
3671 * qpriv->rnr_nak_state is used to determine when the scheduled RNR NAK
3672 * has actually been sent. qp->s_flags RVT_S_ACK_PENDING bit cannot be
3673 * used for this because qp->s_lock is dropped before calling
3674 * hfi1_send_rc_ack() leading to inconsistency between the receive
3675 * interrupt handlers and the send thread in make_rc_ack()
3677 qpriv->rnr_nak_state = TID_RNR_NAK_SEND;
3680 * Schedule RNR NAK to be sent. RNR NAK's are scheduled from the receive
3681 * interrupt handlers but will be sent from the send engine behind any
3682 * previous responses that may have been scheduled
3684 rc_defered_ack(rcd, qp);
3687 void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet)
3689 /* HANDLER FOR TID RDMA WRITE REQUEST packet (Responder side)*/
3692 * 1. Verify TID RDMA WRITE REQ as per IB_OPCODE_RC_RDMA_WRITE_FIRST
3693 * (see hfi1_rc_rcv())
3694 * - Don't allow 0-length requests.
3695 * 2. Put TID RDMA WRITE REQ into the response queueu (s_ack_queue)
3696 * - Setup struct tid_rdma_req with request info
3697 * - Prepare struct tid_rdma_flow array?
3698 * 3. Set the qp->s_ack_state as state diagram in design doc.
3699 * 4. Set RVT_S_RESP_PENDING in s_flags.
3700 * 5. Kick the send engine (hfi1_schedule_send())
3702 struct hfi1_ctxtdata *rcd = packet->rcd;
3703 struct rvt_qp *qp = packet->qp;
3704 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
3705 struct ib_other_headers *ohdr = packet->ohdr;
3706 struct rvt_ack_entry *e;
3707 unsigned long flags;
3708 struct ib_reth *reth;
3709 struct hfi1_qp_priv *qpriv = qp->priv;
3710 struct tid_rdma_request *req;
3711 u32 bth0, psn, len, rkey, num_segs;
3717 bth0 = be32_to_cpu(ohdr->bth[0]);
3718 if (hfi1_ruc_check_hdr(ibp, packet))
3721 fecn = process_ecn(qp, packet);
3722 psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
3723 trace_hfi1_rsp_rcv_tid_write_req(qp, psn);
3725 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
3728 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
3731 reth = &ohdr->u.tid_rdma.w_req.reth;
3732 vaddr = be64_to_cpu(reth->vaddr);
3733 len = be32_to_cpu(reth->length);
3735 num_segs = DIV_ROUND_UP(len, qpriv->tid_rdma.local.max_len);
3736 diff = delta_psn(psn, qp->r_psn);
3737 if (unlikely(diff)) {
3738 tid_rdma_rcv_err(packet, ohdr, qp, psn, diff, fecn);
3743 * The resent request which was previously RNR NAK'd is inserted at the
3744 * location of the original request, which is one entry behind
3747 if (qpriv->rnr_nak_state)
3748 qp->r_head_ack_queue = qp->r_head_ack_queue ?
3749 qp->r_head_ack_queue - 1 :
3750 rvt_size_atomic(ib_to_rvt(qp->ibqp.device));
3752 /* We've verified the request, insert it into the ack queue. */
3753 next = qp->r_head_ack_queue + 1;
3754 if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
3756 spin_lock_irqsave(&qp->s_lock, flags);
3757 if (unlikely(next == qp->s_acked_ack_queue)) {
3758 if (!qp->s_ack_queue[next].sent)
3759 goto nack_inv_unlock;
3760 update_ack_queue(qp, next);
3762 e = &qp->s_ack_queue[qp->r_head_ack_queue];
3763 req = ack_to_tid_req(e);
3765 /* Bring previously RNR NAK'd request back to life */
3766 if (qpriv->rnr_nak_state) {
3767 qp->r_nak_state = 0;
3768 qp->s_nak_state = 0;
3769 qpriv->rnr_nak_state = TID_RNR_NAK_INIT;
3770 qp->r_psn = e->lpsn + 1;
3771 req->state = TID_REQUEST_INIT;
3775 release_rdma_sge_mr(e);
3777 /* The length needs to be in multiples of PAGE_SIZE */
3778 if (!len || len & ~PAGE_MASK)
3779 goto nack_inv_unlock;
3781 rkey = be32_to_cpu(reth->rkey);
3784 if (e->opcode == TID_OP(WRITE_REQ) &&
3785 (req->setup_head != req->clear_tail ||
3786 req->clear_tail != req->acked_tail))
3787 goto nack_inv_unlock;
3789 if (unlikely(!rvt_rkey_ok(qp, &e->rdma_sge, qp->r_len, vaddr,
3790 rkey, IB_ACCESS_REMOTE_WRITE)))
3793 qp->r_psn += num_segs - 1;
3795 e->opcode = (bth0 >> 24) & 0xff;
3797 e->lpsn = qp->r_psn;
3800 req->n_flows = min_t(u16, num_segs, qpriv->tid_rdma.local.max_write);
3801 req->state = TID_REQUEST_INIT;
3807 req->seg_len = qpriv->tid_rdma.local.max_len;
3808 req->total_len = len;
3809 req->total_segs = num_segs;
3810 req->r_flow_psn = e->psn;
3811 req->ss.sge = e->rdma_sge;
3812 req->ss.num_sge = 1;
3814 req->flow_idx = req->setup_head;
3815 req->clear_tail = req->setup_head;
3816 req->acked_tail = req->setup_head;
3818 qp->r_state = e->opcode;
3819 qp->r_nak_state = 0;
3821 * We need to increment the MSN here instead of when we
3822 * finish sending the result since a duplicate request would
3823 * increment it more than once.
3828 trace_hfi1_tid_req_rcv_write_req(qp, 0, e->opcode, e->psn, e->lpsn,
3831 if (qpriv->r_tid_tail == HFI1_QP_WQE_INVALID) {
3832 qpriv->r_tid_tail = qp->r_head_ack_queue;
3833 } else if (qpriv->r_tid_tail == qpriv->r_tid_head) {
3834 struct tid_rdma_request *ptr;
3836 e = &qp->s_ack_queue[qpriv->r_tid_tail];
3837 ptr = ack_to_tid_req(e);
3839 if (e->opcode != TID_OP(WRITE_REQ) ||
3840 ptr->comp_seg == ptr->total_segs) {
3841 if (qpriv->r_tid_tail == qpriv->r_tid_ack)
3842 qpriv->r_tid_ack = qp->r_head_ack_queue;
3843 qpriv->r_tid_tail = qp->r_head_ack_queue;
3847 qp->r_head_ack_queue = next;
3848 qpriv->r_tid_head = qp->r_head_ack_queue;
3850 hfi1_tid_write_alloc_resources(qp, true);
3851 trace_hfi1_tid_write_rsp_rcv_req(qp);
3853 /* Schedule the send tasklet. */
3854 qp->s_flags |= RVT_S_RESP_PENDING;
3856 qp->s_flags |= RVT_S_ECN;
3857 hfi1_schedule_send(qp);
3859 spin_unlock_irqrestore(&qp->s_lock, flags);
3863 spin_unlock_irqrestore(&qp->s_lock, flags);
3865 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
3866 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
3867 qp->r_ack_psn = qp->r_psn;
3868 /* Queue NAK for later */
3869 rc_defered_ack(rcd, qp);
3872 spin_unlock_irqrestore(&qp->s_lock, flags);
3873 rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
3874 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
3875 qp->r_ack_psn = qp->r_psn;
3878 u32 hfi1_build_tid_rdma_write_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,
3879 struct ib_other_headers *ohdr, u32 *bth1,
3881 struct rvt_sge_state **ss)
3883 struct hfi1_ack_priv *epriv = e->priv;
3884 struct tid_rdma_request *req = &epriv->tid_req;
3885 struct hfi1_qp_priv *qpriv = qp->priv;
3886 struct tid_rdma_flow *flow = NULL;
3887 u32 resp_len = 0, hdwords = 0;
3888 void *resp_addr = NULL;
3889 struct tid_rdma_params *remote;
3891 trace_hfi1_tid_req_build_write_resp(qp, 0, e->opcode, e->psn, e->lpsn,
3893 trace_hfi1_tid_write_rsp_build_resp(qp);
3894 trace_hfi1_rsp_build_tid_write_resp(qp, bth2);
3895 flow = &req->flows[req->flow_idx];
3896 switch (req->state) {
3899 * Try to allocate resources here in case QP was queued and was
3900 * later scheduled when resources became available
3902 hfi1_tid_write_alloc_resources(qp, false);
3904 /* We've already sent everything which is ready */
3905 if (req->cur_seg >= req->alloc_seg)
3909 * Resources can be assigned but responses cannot be sent in
3910 * rnr_nak state, till the resent request is received
3912 if (qpriv->rnr_nak_state == TID_RNR_NAK_SENT)
3915 req->state = TID_REQUEST_ACTIVE;
3916 trace_hfi1_tid_flow_build_write_resp(qp, req->flow_idx, flow);
3917 req->flow_idx = CIRC_NEXT(req->flow_idx, MAX_FLOWS);
3918 hfi1_add_tid_reap_timer(qp);
3921 case TID_REQUEST_RESEND_ACTIVE:
3922 case TID_REQUEST_RESEND:
3923 trace_hfi1_tid_flow_build_write_resp(qp, req->flow_idx, flow);
3924 req->flow_idx = CIRC_NEXT(req->flow_idx, MAX_FLOWS);
3925 if (!CIRC_CNT(req->setup_head, req->flow_idx, MAX_FLOWS))
3926 req->state = TID_REQUEST_ACTIVE;
3928 hfi1_mod_tid_reap_timer(qp);
3931 flow->flow_state.resp_ib_psn = bth2;
3932 resp_addr = (void *)flow->tid_entry;
3933 resp_len = sizeof(*flow->tid_entry) * flow->tidcnt;
3936 memset(&ohdr->u.tid_rdma.w_rsp, 0, sizeof(ohdr->u.tid_rdma.w_rsp));
3937 epriv->ss.sge.vaddr = resp_addr;
3938 epriv->ss.sge.sge_length = resp_len;
3939 epriv->ss.sge.length = epriv->ss.sge.sge_length;
3941 * We can safely zero these out. Since the first SGE covers the
3942 * entire packet, nothing else should even look at the MR.
3944 epriv->ss.sge.mr = NULL;
3945 epriv->ss.sge.m = 0;
3946 epriv->ss.sge.n = 0;
3948 epriv->ss.sg_list = NULL;
3949 epriv->ss.total_len = epriv->ss.sge.sge_length;
3950 epriv->ss.num_sge = 1;
3953 *len = epriv->ss.total_len;
3955 /* Construct the TID RDMA WRITE RESP packet header */
3957 remote = rcu_dereference(qpriv->tid_rdma.remote);
3959 KDETH_RESET(ohdr->u.tid_rdma.w_rsp.kdeth0, KVER, 0x1);
3960 KDETH_RESET(ohdr->u.tid_rdma.w_rsp.kdeth1, JKEY, remote->jkey);
3961 ohdr->u.tid_rdma.w_rsp.aeth = rvt_compute_aeth(qp);
3962 ohdr->u.tid_rdma.w_rsp.tid_flow_psn =
3963 cpu_to_be32((flow->flow_state.generation <<
3964 HFI1_KDETH_BTH_SEQ_SHIFT) |
3965 (flow->flow_state.spsn &
3966 HFI1_KDETH_BTH_SEQ_MASK));
3967 ohdr->u.tid_rdma.w_rsp.tid_flow_qp =
3968 cpu_to_be32(qpriv->tid_rdma.local.qp |
3969 ((flow->idx & TID_RDMA_DESTQP_FLOW_MASK) <<
3970 TID_RDMA_DESTQP_FLOW_SHIFT) |
3972 ohdr->u.tid_rdma.w_rsp.verbs_qp = cpu_to_be32(qp->remote_qpn);
3975 hdwords = sizeof(ohdr->u.tid_rdma.w_rsp) / sizeof(u32);
3976 qpriv->pending_tid_w_segs++;
3981 static void hfi1_add_tid_reap_timer(struct rvt_qp *qp)
3983 struct hfi1_qp_priv *qpriv = qp->priv;
3985 lockdep_assert_held(&qp->s_lock);
3986 if (!(qpriv->s_flags & HFI1_R_TID_RSC_TIMER)) {
3987 qpriv->s_flags |= HFI1_R_TID_RSC_TIMER;
3988 qpriv->s_tid_timer.expires = jiffies +
3989 qpriv->tid_timer_timeout_jiffies;
3990 add_timer(&qpriv->s_tid_timer);
3994 static void hfi1_mod_tid_reap_timer(struct rvt_qp *qp)
3996 struct hfi1_qp_priv *qpriv = qp->priv;
3998 lockdep_assert_held(&qp->s_lock);
3999 qpriv->s_flags |= HFI1_R_TID_RSC_TIMER;
4000 mod_timer(&qpriv->s_tid_timer, jiffies +
4001 qpriv->tid_timer_timeout_jiffies);
4004 static int hfi1_stop_tid_reap_timer(struct rvt_qp *qp)
4006 struct hfi1_qp_priv *qpriv = qp->priv;
4009 lockdep_assert_held(&qp->s_lock);
4010 if (qpriv->s_flags & HFI1_R_TID_RSC_TIMER) {
4011 rval = del_timer(&qpriv->s_tid_timer);
4012 qpriv->s_flags &= ~HFI1_R_TID_RSC_TIMER;
4017 void hfi1_del_tid_reap_timer(struct rvt_qp *qp)
4019 struct hfi1_qp_priv *qpriv = qp->priv;
4021 del_timer_sync(&qpriv->s_tid_timer);
4022 qpriv->s_flags &= ~HFI1_R_TID_RSC_TIMER;
4025 static void hfi1_tid_timeout(struct timer_list *t)
4027 struct hfi1_qp_priv *qpriv = from_timer(qpriv, t, s_tid_timer);
4028 struct rvt_qp *qp = qpriv->owner;
4029 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
4030 unsigned long flags;
4033 spin_lock_irqsave(&qp->r_lock, flags);
4034 spin_lock(&qp->s_lock);
4035 if (qpriv->s_flags & HFI1_R_TID_RSC_TIMER) {
4036 dd_dev_warn(dd_from_ibdev(qp->ibqp.device), "[QP%u] %s %d\n",
4037 qp->ibqp.qp_num, __func__, __LINE__);
4038 trace_hfi1_msg_tid_timeout(/* msg */
4039 qp, "resource timeout = ",
4040 (u64)qpriv->tid_timer_timeout_jiffies);
4041 hfi1_stop_tid_reap_timer(qp);
4043 * Go though the entire ack queue and clear any outstanding
4044 * HW flow and RcvArray resources.
4046 hfi1_kern_clear_hw_flow(qpriv->rcd, qp);
4047 for (i = 0; i < rvt_max_atomic(rdi); i++) {
4048 struct tid_rdma_request *req =
4049 ack_to_tid_req(&qp->s_ack_queue[i]);
4051 hfi1_kern_exp_rcv_clear_all(req);
4053 spin_unlock(&qp->s_lock);
4054 if (qp->ibqp.event_handler) {
4057 ev.device = qp->ibqp.device;
4058 ev.element.qp = &qp->ibqp;
4059 ev.event = IB_EVENT_QP_FATAL;
4060 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
4062 rvt_rc_error(qp, IB_WC_RESP_TIMEOUT_ERR);
4065 spin_unlock(&qp->s_lock);
4067 spin_unlock_irqrestore(&qp->r_lock, flags);
4070 void hfi1_rc_rcv_tid_rdma_write_resp(struct hfi1_packet *packet)
4072 /* HANDLER FOR TID RDMA WRITE RESPONSE packet (Requestor side */
4075 * 1. Find matching SWQE
4076 * 2. Check that TIDENTRY array has enough space for a complete
4077 * segment. If not, put QP in error state.
4078 * 3. Save response data in struct tid_rdma_req and struct tid_rdma_flow
4079 * 4. Remove HFI1_S_WAIT_TID_RESP from s_flags.
4080 * 5. Set qp->s_state
4081 * 6. Kick the send engine (hfi1_schedule_send())
4083 struct ib_other_headers *ohdr = packet->ohdr;
4084 struct rvt_qp *qp = packet->qp;
4085 struct hfi1_qp_priv *qpriv = qp->priv;
4086 struct hfi1_ctxtdata *rcd = packet->rcd;
4087 struct rvt_swqe *wqe;
4088 struct tid_rdma_request *req;
4089 struct tid_rdma_flow *flow;
4090 enum ib_wc_status status;
4091 u32 opcode, aeth, psn, flow_psn, i, tidlen = 0, pktlen;
4093 unsigned long flags;
4095 fecn = process_ecn(qp, packet);
4096 psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
4097 aeth = be32_to_cpu(ohdr->u.tid_rdma.w_rsp.aeth);
4098 opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff;
4100 spin_lock_irqsave(&qp->s_lock, flags);
4102 /* Ignore invalid responses */
4103 if (cmp_psn(psn, qp->s_next_psn) >= 0)
4106 /* Ignore duplicate responses. */
4107 if (unlikely(cmp_psn(psn, qp->s_last_psn) <= 0))
4110 if (unlikely(qp->s_acked == qp->s_tail))
4114 * If we are waiting for a particular packet sequence number
4115 * due to a request being resent, check for it. Otherwise,
4116 * ensure that we haven't missed anything.
4118 if (qp->r_flags & RVT_R_RDMAR_SEQ) {
4119 if (cmp_psn(psn, qp->s_last_psn + 1) != 0)
4121 qp->r_flags &= ~RVT_R_RDMAR_SEQ;
4124 wqe = rvt_get_swqe_ptr(qp, qpriv->s_tid_cur);
4125 if (unlikely(wqe->wr.opcode != IB_WR_TID_RDMA_WRITE))
4128 req = wqe_to_tid_req(wqe);
4130 * If we've lost ACKs and our acked_tail pointer is too far
4131 * behind, don't overwrite segments. Just drop the packet and
4132 * let the reliability protocol take care of it.
4134 if (!CIRC_SPACE(req->setup_head, req->acked_tail, MAX_FLOWS))
4138 * The call to do_rc_ack() should be last in the chain of
4139 * packet checks because it will end up updating the QP state.
4140 * Therefore, anything that would prevent the packet from
4141 * being accepted as a successful response should be prior
4144 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
4147 trace_hfi1_ack(qp, psn);
4149 flow = &req->flows[req->setup_head];
4152 flow->tid_offset = 0;
4154 flow->resync_npkts = 0;
4155 flow->tid_qpn = be32_to_cpu(ohdr->u.tid_rdma.w_rsp.tid_flow_qp);
4156 flow->idx = (flow->tid_qpn >> TID_RDMA_DESTQP_FLOW_SHIFT) &
4157 TID_RDMA_DESTQP_FLOW_MASK;
4158 flow_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.w_rsp.tid_flow_psn));
4159 flow->flow_state.generation = flow_psn >> HFI1_KDETH_BTH_SEQ_SHIFT;
4160 flow->flow_state.spsn = flow_psn & HFI1_KDETH_BTH_SEQ_MASK;
4161 flow->flow_state.resp_ib_psn = psn;
4162 flow->length = min_t(u32, req->seg_len,
4163 (wqe->length - (req->comp_seg * req->seg_len)));
4165 flow->npkts = rvt_div_round_up_mtu(qp, flow->length);
4166 flow->flow_state.lpsn = flow->flow_state.spsn +
4168 /* payload length = packet length - (header length + ICRC length) */
4169 pktlen = packet->tlen - (packet->hlen + 4);
4170 if (pktlen > sizeof(flow->tid_entry)) {
4171 status = IB_WC_LOC_LEN_ERR;
4174 memcpy(flow->tid_entry, packet->ebuf, pktlen);
4175 flow->tidcnt = pktlen / sizeof(*flow->tid_entry);
4176 trace_hfi1_tid_flow_rcv_write_resp(qp, req->setup_head, flow);
4179 trace_hfi1_tid_write_sender_rcv_resp(qp, 0);
4181 * Walk the TID_ENTRY list to make sure we have enough space for a
4184 for (i = 0; i < flow->tidcnt; i++) {
4185 trace_hfi1_tid_entry_rcv_write_resp(/* entry */
4186 qp, i, flow->tid_entry[i]);
4187 if (!EXP_TID_GET(flow->tid_entry[i], LEN)) {
4188 status = IB_WC_LOC_LEN_ERR;
4191 tidlen += EXP_TID_GET(flow->tid_entry[i], LEN);
4193 if (tidlen * PAGE_SIZE < flow->length) {
4194 status = IB_WC_LOC_LEN_ERR;
4198 trace_hfi1_tid_req_rcv_write_resp(qp, 0, wqe->wr.opcode, wqe->psn,
4201 * If this is the first response for this request, set the initial
4202 * flow index to the current flow.
4204 if (!cmp_psn(psn, wqe->psn)) {
4205 req->r_last_acked = mask_psn(wqe->psn - 1);
4206 /* Set acked flow index to head index */
4207 req->acked_tail = req->setup_head;
4210 /* advance circular buffer head */
4211 req->setup_head = CIRC_NEXT(req->setup_head, MAX_FLOWS);
4212 req->state = TID_REQUEST_ACTIVE;
4215 * If all responses for this TID RDMA WRITE request have been received
4216 * advance the pointer to the next one.
4217 * Since TID RDMA requests could be mixed in with regular IB requests,
4218 * they might not appear sequentially in the queue. Therefore, the
4219 * next request needs to be "found".
4221 if (qpriv->s_tid_cur != qpriv->s_tid_head &&
4222 req->comp_seg == req->total_segs) {
4223 for (i = qpriv->s_tid_cur + 1; ; i++) {
4224 if (i == qp->s_size)
4226 wqe = rvt_get_swqe_ptr(qp, i);
4227 if (i == qpriv->s_tid_head)
4229 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE)
4232 qpriv->s_tid_cur = i;
4234 qp->s_flags &= ~HFI1_S_WAIT_TID_RESP;
4235 hfi1_schedule_tid_send(qp);
4239 status = IB_WC_LOC_QP_OP_ERR;
4241 rvt_error_qp(qp, status);
4244 qp->s_flags |= RVT_S_ECN;
4245 spin_unlock_irqrestore(&qp->s_lock, flags);
4248 bool hfi1_build_tid_rdma_packet(struct rvt_swqe *wqe,
4249 struct ib_other_headers *ohdr,
4250 u32 *bth1, u32 *bth2, u32 *len)
4252 struct tid_rdma_request *req = wqe_to_tid_req(wqe);
4253 struct tid_rdma_flow *flow = &req->flows[req->clear_tail];
4254 struct tid_rdma_params *remote;
4255 struct rvt_qp *qp = req->qp;
4256 struct hfi1_qp_priv *qpriv = qp->priv;
4257 u32 tidentry = flow->tid_entry[flow->tid_idx];
4258 u32 tidlen = EXP_TID_GET(tidentry, LEN) << PAGE_SHIFT;
4259 struct tid_rdma_write_data *wd = &ohdr->u.tid_rdma.w_data;
4260 u32 next_offset, om = KDETH_OM_LARGE;
4264 hfi1_trdma_send_complete(qp, wqe, IB_WC_REM_INV_RD_REQ_ERR);
4265 rvt_error_qp(qp, IB_WC_REM_INV_RD_REQ_ERR);
4268 *len = min_t(u32, qp->pmtu, tidlen - flow->tid_offset);
4270 next_offset = flow->tid_offset + *len;
4271 last_pkt = (flow->tid_idx == (flow->tidcnt - 1) &&
4272 next_offset >= tidlen) || (flow->sent >= flow->length);
4273 trace_hfi1_tid_entry_build_write_data(qp, flow->tid_idx, tidentry);
4274 trace_hfi1_tid_flow_build_write_data(qp, req->clear_tail, flow);
4277 remote = rcu_dereference(qpriv->tid_rdma.remote);
4278 KDETH_RESET(wd->kdeth0, KVER, 0x1);
4279 KDETH_SET(wd->kdeth0, SH, !last_pkt);
4280 KDETH_SET(wd->kdeth0, INTR, !!(!last_pkt && remote->urg));
4281 KDETH_SET(wd->kdeth0, TIDCTRL, EXP_TID_GET(tidentry, CTRL));
4282 KDETH_SET(wd->kdeth0, TID, EXP_TID_GET(tidentry, IDX));
4283 KDETH_SET(wd->kdeth0, OM, om == KDETH_OM_LARGE);
4284 KDETH_SET(wd->kdeth0, OFFSET, flow->tid_offset / om);
4285 KDETH_RESET(wd->kdeth1, JKEY, remote->jkey);
4286 wd->verbs_qp = cpu_to_be32(qp->remote_qpn);
4289 *bth1 = flow->tid_qpn;
4290 *bth2 = mask_psn(((flow->flow_state.spsn + flow->pkt++) &
4291 HFI1_KDETH_BTH_SEQ_MASK) |
4292 (flow->flow_state.generation <<
4293 HFI1_KDETH_BTH_SEQ_SHIFT));
4295 /* PSNs are zero-based, so +1 to count number of packets */
4296 if (flow->flow_state.lpsn + 1 +
4297 rvt_div_round_up_mtu(qp, req->seg_len) >
4299 req->state = TID_REQUEST_SYNC;
4300 *bth2 |= IB_BTH_REQ_ACK;
4303 if (next_offset >= tidlen) {
4304 flow->tid_offset = 0;
4307 flow->tid_offset = next_offset;
4312 void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet)
4314 struct rvt_qp *qp = packet->qp;
4315 struct hfi1_qp_priv *priv = qp->priv;
4316 struct hfi1_ctxtdata *rcd = priv->rcd;
4317 struct ib_other_headers *ohdr = packet->ohdr;
4318 struct rvt_ack_entry *e;
4319 struct tid_rdma_request *req;
4320 struct tid_rdma_flow *flow;
4321 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
4322 unsigned long flags;
4327 fecn = process_ecn(qp, packet);
4328 psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
4329 opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff;
4332 * All error handling should be done by now. If we are here, the packet
4333 * is either good or been accepted by the error handler.
4335 spin_lock_irqsave(&qp->s_lock, flags);
4336 e = &qp->s_ack_queue[priv->r_tid_tail];
4337 req = ack_to_tid_req(e);
4338 flow = &req->flows[req->clear_tail];
4339 if (cmp_psn(psn, full_flow_psn(flow, flow->flow_state.lpsn))) {
4340 update_r_next_psn_fecn(packet, priv, rcd, flow, fecn);
4342 if (cmp_psn(psn, flow->flow_state.r_next_psn))
4345 flow->flow_state.r_next_psn = mask_psn(psn + 1);
4347 * Copy the payload to destination buffer if this packet is
4348 * delivered as an eager packet due to RSM rule and FECN.
4349 * The RSM rule selects FECN bit in BTH and SH bit in
4350 * KDETH header and therefore will not match the last
4351 * packet of each segment that has SH bit cleared.
4353 if (fecn && packet->etype == RHF_RCV_TYPE_EAGER) {
4354 struct rvt_sge_state ss;
4356 u32 tlen = packet->tlen;
4357 u16 hdrsize = packet->hlen;
4358 u8 pad = packet->pad;
4359 u8 extra_bytes = pad + packet->extra_byte +
4361 u32 pmtu = qp->pmtu;
4363 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
4365 len = req->comp_seg * req->seg_len;
4366 len += delta_psn(psn,
4367 full_flow_psn(flow, flow->flow_state.spsn)) *
4369 if (unlikely(req->total_len - len < pmtu))
4373 * The e->rdma_sge field is set when TID RDMA WRITE REQ
4374 * is first received and is never modified thereafter.
4376 ss.sge = e->rdma_sge;
4379 ss.total_len = req->total_len;
4380 rvt_skip_sge(&ss, len, false);
4381 rvt_copy_sge(qp, &ss, packet->payload, pmtu, false,
4383 /* Raise the sw sequence check flag for next packet */
4384 priv->r_next_psn_kdeth = mask_psn(psn + 1);
4385 priv->s_flags |= HFI1_R_TID_SW_PSN;
4389 flow->flow_state.r_next_psn = mask_psn(psn + 1);
4390 hfi1_kern_exp_rcv_clear(req);
4391 priv->alloc_w_segs--;
4392 rcd->flows[flow->idx].psn = psn & HFI1_KDETH_BTH_SEQ_MASK;
4394 priv->s_nak_state = 0;
4397 * Release the flow if one of the following conditions has been met:
4398 * - The request has reached a sync point AND all outstanding
4399 * segments have been completed, or
4400 * - The entire request is complete and there are no more requests
4401 * (of any kind) in the queue.
4403 trace_hfi1_rsp_rcv_tid_write_data(qp, psn);
4404 trace_hfi1_tid_req_rcv_write_data(qp, 0, e->opcode, e->psn, e->lpsn,
4406 trace_hfi1_tid_write_rsp_rcv_data(qp);
4407 if (priv->r_tid_ack == HFI1_QP_WQE_INVALID)
4408 priv->r_tid_ack = priv->r_tid_tail;
4410 if (opcode == TID_OP(WRITE_DATA_LAST)) {
4411 release_rdma_sge_mr(e);
4412 for (next = priv->r_tid_tail + 1; ; next++) {
4413 if (next > rvt_size_atomic(&dev->rdi))
4415 if (next == priv->r_tid_head)
4417 e = &qp->s_ack_queue[next];
4418 if (e->opcode == TID_OP(WRITE_REQ))
4421 priv->r_tid_tail = next;
4422 if (++qp->s_acked_ack_queue > rvt_size_atomic(&dev->rdi))
4423 qp->s_acked_ack_queue = 0;
4426 hfi1_tid_write_alloc_resources(qp, true);
4429 * If we need to generate more responses, schedule the
4432 if (req->cur_seg < req->total_segs ||
4433 qp->s_tail_ack_queue != qp->r_head_ack_queue) {
4434 qp->s_flags |= RVT_S_RESP_PENDING;
4435 hfi1_schedule_send(qp);
4438 priv->pending_tid_w_segs--;
4439 if (priv->s_flags & HFI1_R_TID_RSC_TIMER) {
4440 if (priv->pending_tid_w_segs)
4441 hfi1_mod_tid_reap_timer(req->qp);
4443 hfi1_stop_tid_reap_timer(req->qp);
4447 priv->s_flags |= RVT_S_ACK_PENDING;
4448 hfi1_schedule_tid_send(qp);
4450 priv->r_next_psn_kdeth = flow->flow_state.r_next_psn;
4452 qp->s_flags |= RVT_S_ECN;
4453 spin_unlock_irqrestore(&qp->s_lock, flags);
4457 if (!priv->s_nak_state) {
4458 priv->s_nak_state = IB_NAK_PSN_ERROR;
4459 priv->s_nak_psn = flow->flow_state.r_next_psn;
4460 priv->s_flags |= RVT_S_ACK_PENDING;
4461 if (priv->r_tid_ack == HFI1_QP_WQE_INVALID)
4462 priv->r_tid_ack = priv->r_tid_tail;
4463 hfi1_schedule_tid_send(qp);
4468 static bool hfi1_tid_rdma_is_resync_psn(u32 psn)
4470 return (bool)((psn & HFI1_KDETH_BTH_SEQ_MASK) ==
4471 HFI1_KDETH_BTH_SEQ_MASK);
4474 u32 hfi1_build_tid_rdma_write_ack(struct rvt_qp *qp, struct rvt_ack_entry *e,
4475 struct ib_other_headers *ohdr, u16 iflow,
4476 u32 *bth1, u32 *bth2)
4478 struct hfi1_qp_priv *qpriv = qp->priv;
4479 struct tid_flow_state *fs = &qpriv->flow_state;
4480 struct tid_rdma_request *req = ack_to_tid_req(e);
4481 struct tid_rdma_flow *flow = &req->flows[iflow];
4482 struct tid_rdma_params *remote;
4485 remote = rcu_dereference(qpriv->tid_rdma.remote);
4486 KDETH_RESET(ohdr->u.tid_rdma.ack.kdeth1, JKEY, remote->jkey);
4487 ohdr->u.tid_rdma.ack.verbs_qp = cpu_to_be32(qp->remote_qpn);
4491 if (qpriv->resync) {
4492 *bth2 = mask_psn((fs->generation <<
4493 HFI1_KDETH_BTH_SEQ_SHIFT) - 1);
4494 ohdr->u.tid_rdma.ack.aeth = rvt_compute_aeth(qp);
4495 } else if (qpriv->s_nak_state) {
4496 *bth2 = mask_psn(qpriv->s_nak_psn);
4497 ohdr->u.tid_rdma.ack.aeth =
4498 cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
4499 (qpriv->s_nak_state <<
4500 IB_AETH_CREDIT_SHIFT));
4502 *bth2 = full_flow_psn(flow, flow->flow_state.lpsn);
4503 ohdr->u.tid_rdma.ack.aeth = rvt_compute_aeth(qp);
4505 KDETH_RESET(ohdr->u.tid_rdma.ack.kdeth0, KVER, 0x1);
4506 ohdr->u.tid_rdma.ack.tid_flow_qp =
4507 cpu_to_be32(qpriv->tid_rdma.local.qp |
4508 ((flow->idx & TID_RDMA_DESTQP_FLOW_MASK) <<
4509 TID_RDMA_DESTQP_FLOW_SHIFT) |
4512 ohdr->u.tid_rdma.ack.tid_flow_psn = 0;
4513 ohdr->u.tid_rdma.ack.verbs_psn =
4514 cpu_to_be32(flow->flow_state.resp_ib_psn);
4516 if (qpriv->resync) {
4518 * If the PSN before the current expect KDETH PSN is the
4519 * RESYNC PSN, then we never received a good TID RDMA WRITE
4520 * DATA packet after a previous RESYNC.
4521 * In this case, the next expected KDETH PSN stays the same.
4523 if (hfi1_tid_rdma_is_resync_psn(qpriv->r_next_psn_kdeth - 1)) {
4524 ohdr->u.tid_rdma.ack.tid_flow_psn =
4525 cpu_to_be32(qpriv->r_next_psn_kdeth_save);
4528 * Because the KDETH PSNs jump during a RESYNC, it's
4529 * not possible to infer (or compute) the previous value
4530 * of r_next_psn_kdeth in the case of back-to-back
4531 * RESYNC packets. Therefore, we save it.
4533 qpriv->r_next_psn_kdeth_save =
4534 qpriv->r_next_psn_kdeth - 1;
4535 ohdr->u.tid_rdma.ack.tid_flow_psn =
4536 cpu_to_be32(qpriv->r_next_psn_kdeth_save);
4537 qpriv->r_next_psn_kdeth = mask_psn(*bth2 + 1);
4539 qpriv->resync = false;
4542 return sizeof(ohdr->u.tid_rdma.ack) / sizeof(u32);
4545 void hfi1_rc_rcv_tid_rdma_ack(struct hfi1_packet *packet)
4547 struct ib_other_headers *ohdr = packet->ohdr;
4548 struct rvt_qp *qp = packet->qp;
4549 struct hfi1_qp_priv *qpriv = qp->priv;
4550 struct rvt_swqe *wqe;
4551 struct tid_rdma_request *req;
4552 struct tid_rdma_flow *flow;
4553 u32 aeth, psn, req_psn, ack_psn, fspsn, resync_psn, ack_kpsn;
4554 unsigned long flags;
4557 trace_hfi1_tid_write_sender_rcv_tid_ack(qp, 0);
4558 process_ecn(qp, packet);
4559 psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
4560 aeth = be32_to_cpu(ohdr->u.tid_rdma.ack.aeth);
4561 req_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.ack.verbs_psn));
4562 resync_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.ack.tid_flow_psn));
4564 spin_lock_irqsave(&qp->s_lock, flags);
4565 trace_hfi1_rcv_tid_ack(qp, aeth, psn, req_psn, resync_psn);
4567 /* If we are waiting for an ACK to RESYNC, drop any other packets */
4568 if ((qp->s_flags & HFI1_S_WAIT_HALT) &&
4569 cmp_psn(psn, qpriv->s_resync_psn))
4573 if (hfi1_tid_rdma_is_resync_psn(psn))
4574 ack_kpsn = resync_psn;
4582 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
4584 if (wqe->wr.opcode != IB_WR_TID_RDMA_WRITE)
4587 req = wqe_to_tid_req(wqe);
4588 trace_hfi1_tid_req_rcv_tid_ack(qp, 0, wqe->wr.opcode, wqe->psn,
4590 flow = &req->flows[req->acked_tail];
4591 trace_hfi1_tid_flow_rcv_tid_ack(qp, req->acked_tail, flow);
4593 /* Drop stale ACK/NAK */
4594 if (cmp_psn(psn, full_flow_psn(flow, flow->flow_state.spsn)) < 0)
4597 while (cmp_psn(ack_kpsn,
4598 full_flow_psn(flow, flow->flow_state.lpsn)) >= 0 &&
4599 req->ack_seg < req->cur_seg) {
4601 /* advance acked segment pointer */
4602 req->acked_tail = CIRC_NEXT(req->acked_tail, MAX_FLOWS);
4603 req->r_last_acked = flow->flow_state.resp_ib_psn;
4604 trace_hfi1_tid_req_rcv_tid_ack(qp, 0, wqe->wr.opcode, wqe->psn,
4606 if (req->ack_seg == req->total_segs) {
4607 req->state = TID_REQUEST_COMPLETE;
4608 wqe = do_rc_completion(qp, wqe,
4609 to_iport(qp->ibqp.device,
4611 trace_hfi1_sender_rcv_tid_ack(qp);
4612 atomic_dec(&qpriv->n_tid_requests);
4613 if (qp->s_acked == qp->s_tail)
4615 if (wqe->wr.opcode != IB_WR_TID_RDMA_WRITE)
4617 req = wqe_to_tid_req(wqe);
4619 flow = &req->flows[req->acked_tail];
4620 trace_hfi1_tid_flow_rcv_tid_ack(qp, req->acked_tail, flow);
4623 trace_hfi1_tid_req_rcv_tid_ack(qp, 0, wqe->wr.opcode, wqe->psn,
4625 switch (aeth >> 29) {
4627 if (qpriv->s_flags & RVT_S_WAIT_ACK)
4628 qpriv->s_flags &= ~RVT_S_WAIT_ACK;
4629 if (!hfi1_tid_rdma_is_resync_psn(psn)) {
4630 /* Check if there is any pending TID ACK */
4631 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE &&
4632 req->ack_seg < req->cur_seg)
4633 hfi1_mod_tid_retry_timer(qp);
4635 hfi1_stop_tid_retry_timer(qp);
4636 hfi1_schedule_send(qp);
4638 u32 spsn, fpsn, last_acked, generation;
4639 struct tid_rdma_request *rptr;
4642 hfi1_stop_tid_retry_timer(qp);
4643 /* Allow new requests (see hfi1_make_tid_rdma_pkt) */
4644 qp->s_flags &= ~HFI1_S_WAIT_HALT;
4646 * Clear RVT_S_SEND_ONE flag in case that the TID RDMA
4647 * ACK is received after the TID retry timer is fired
4648 * again. In this case, do not send any more TID
4649 * RESYNC request or wait for any more TID ACK packet.
4651 qpriv->s_flags &= ~RVT_S_SEND_ONE;
4652 hfi1_schedule_send(qp);
4654 if ((qp->s_acked == qpriv->s_tid_tail &&
4655 req->ack_seg == req->total_segs) ||
4656 qp->s_acked == qp->s_tail) {
4657 qpriv->s_state = TID_OP(WRITE_DATA_LAST);
4661 if (req->ack_seg == req->comp_seg) {
4662 qpriv->s_state = TID_OP(WRITE_DATA);
4667 * The PSN to start with is the next PSN after the
4670 psn = mask_psn(psn + 1);
4671 generation = psn >> HFI1_KDETH_BTH_SEQ_SHIFT;
4675 * Update to the correct WQE when we get an ACK(RESYNC)
4676 * in the middle of a request.
4678 if (delta_psn(ack_psn, wqe->lpsn))
4679 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
4680 req = wqe_to_tid_req(wqe);
4681 flow = &req->flows[req->acked_tail];
4683 * RESYNC re-numbers the PSN ranges of all remaining
4684 * segments. Also, PSN's start from 0 in the middle of a
4685 * segment and the first segment size is less than the
4686 * default number of packets. flow->resync_npkts is used
4687 * to track the number of packets from the start of the
4688 * real segment to the point of 0 PSN after the RESYNC
4689 * in order to later correctly rewind the SGE.
4691 fpsn = full_flow_psn(flow, flow->flow_state.spsn);
4692 req->r_ack_psn = psn;
4693 flow->resync_npkts +=
4694 delta_psn(mask_psn(resync_psn + 1), fpsn);
4696 * Renumber all packet sequence number ranges
4697 * based on the new generation.
4699 last_acked = qp->s_acked;
4702 /* start from last acked segment */
4703 for (fidx = rptr->acked_tail;
4704 CIRC_CNT(rptr->setup_head, fidx,
4706 fidx = CIRC_NEXT(fidx, MAX_FLOWS)) {
4710 flow = &rptr->flows[fidx];
4711 gen = flow->flow_state.generation;
4712 if (WARN_ON(gen == generation &&
4713 flow->flow_state.spsn !=
4716 lpsn = flow->flow_state.lpsn;
4717 lpsn = full_flow_psn(flow, lpsn);
4720 mask_psn(resync_psn)
4722 flow->flow_state.generation =
4724 flow->flow_state.spsn = spsn;
4725 flow->flow_state.lpsn =
4726 flow->flow_state.spsn +
4729 spsn += flow->npkts;
4730 resync_psn += flow->npkts;
4731 trace_hfi1_tid_flow_rcv_tid_ack(qp,
4735 if (++last_acked == qpriv->s_tid_cur + 1)
4737 if (last_acked == qp->s_size)
4739 wqe = rvt_get_swqe_ptr(qp, last_acked);
4740 rptr = wqe_to_tid_req(wqe);
4742 req->cur_seg = req->ack_seg;
4743 qpriv->s_tid_tail = qp->s_acked;
4744 qpriv->s_state = TID_OP(WRITE_REQ);
4745 hfi1_schedule_tid_send(qp);
4748 qpriv->s_retry = qp->s_retry_cnt;
4752 hfi1_stop_tid_retry_timer(qp);
4753 switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
4754 IB_AETH_CREDIT_MASK) {
4755 case 0: /* PSN sequence error */
4756 flow = &req->flows[req->acked_tail];
4757 fspsn = full_flow_psn(flow, flow->flow_state.spsn);
4758 trace_hfi1_tid_flow_rcv_tid_ack(qp, req->acked_tail,
4760 req->r_ack_psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
4761 req->cur_seg = req->ack_seg;
4762 qpriv->s_tid_tail = qp->s_acked;
4763 qpriv->s_state = TID_OP(WRITE_REQ);
4764 qpriv->s_retry = qp->s_retry_cnt;
4765 hfi1_schedule_tid_send(qp);
4778 spin_unlock_irqrestore(&qp->s_lock, flags);
4781 void hfi1_add_tid_retry_timer(struct rvt_qp *qp)
4783 struct hfi1_qp_priv *priv = qp->priv;
4784 struct ib_qp *ibqp = &qp->ibqp;
4785 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
4787 lockdep_assert_held(&qp->s_lock);
4788 if (!(priv->s_flags & HFI1_S_TID_RETRY_TIMER)) {
4789 priv->s_flags |= HFI1_S_TID_RETRY_TIMER;
4790 priv->s_tid_retry_timer.expires = jiffies +
4791 priv->tid_retry_timeout_jiffies + rdi->busy_jiffies;
4792 add_timer(&priv->s_tid_retry_timer);
4796 static void hfi1_mod_tid_retry_timer(struct rvt_qp *qp)
4798 struct hfi1_qp_priv *priv = qp->priv;
4799 struct ib_qp *ibqp = &qp->ibqp;
4800 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
4802 lockdep_assert_held(&qp->s_lock);
4803 priv->s_flags |= HFI1_S_TID_RETRY_TIMER;
4804 mod_timer(&priv->s_tid_retry_timer, jiffies +
4805 priv->tid_retry_timeout_jiffies + rdi->busy_jiffies);
4808 static int hfi1_stop_tid_retry_timer(struct rvt_qp *qp)
4810 struct hfi1_qp_priv *priv = qp->priv;
4813 lockdep_assert_held(&qp->s_lock);
4814 if (priv->s_flags & HFI1_S_TID_RETRY_TIMER) {
4815 rval = del_timer(&priv->s_tid_retry_timer);
4816 priv->s_flags &= ~HFI1_S_TID_RETRY_TIMER;
4821 void hfi1_del_tid_retry_timer(struct rvt_qp *qp)
4823 struct hfi1_qp_priv *priv = qp->priv;
4825 del_timer_sync(&priv->s_tid_retry_timer);
4826 priv->s_flags &= ~HFI1_S_TID_RETRY_TIMER;
4829 static void hfi1_tid_retry_timeout(struct timer_list *t)
4831 struct hfi1_qp_priv *priv = from_timer(priv, t, s_tid_retry_timer);
4832 struct rvt_qp *qp = priv->owner;
4833 struct rvt_swqe *wqe;
4834 unsigned long flags;
4835 struct tid_rdma_request *req;
4837 spin_lock_irqsave(&qp->r_lock, flags);
4838 spin_lock(&qp->s_lock);
4839 trace_hfi1_tid_write_sender_retry_timeout(qp, 0);
4840 if (priv->s_flags & HFI1_S_TID_RETRY_TIMER) {
4841 hfi1_stop_tid_retry_timer(qp);
4842 if (!priv->s_retry) {
4843 trace_hfi1_msg_tid_retry_timeout(/* msg */
4845 "Exhausted retries. Tid retry timeout = ",
4846 (u64)priv->tid_retry_timeout_jiffies);
4848 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
4849 hfi1_trdma_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
4850 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
4852 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
4853 req = wqe_to_tid_req(wqe);
4854 trace_hfi1_tid_req_tid_retry_timeout(/* req */
4855 qp, 0, wqe->wr.opcode, wqe->psn, wqe->lpsn, req);
4857 priv->s_flags &= ~RVT_S_WAIT_ACK;
4858 /* Only send one packet (the RESYNC) */
4859 priv->s_flags |= RVT_S_SEND_ONE;
4861 * No additional request shall be made by this QP until
4862 * the RESYNC has been complete.
4864 qp->s_flags |= HFI1_S_WAIT_HALT;
4865 priv->s_state = TID_OP(RESYNC);
4867 hfi1_schedule_tid_send(qp);
4870 spin_unlock(&qp->s_lock);
4871 spin_unlock_irqrestore(&qp->r_lock, flags);
4874 u32 hfi1_build_tid_rdma_resync(struct rvt_qp *qp, struct rvt_swqe *wqe,
4875 struct ib_other_headers *ohdr, u32 *bth1,
4876 u32 *bth2, u16 fidx)
4878 struct hfi1_qp_priv *qpriv = qp->priv;
4879 struct tid_rdma_params *remote;
4880 struct tid_rdma_request *req = wqe_to_tid_req(wqe);
4881 struct tid_rdma_flow *flow = &req->flows[fidx];
4885 remote = rcu_dereference(qpriv->tid_rdma.remote);
4886 KDETH_RESET(ohdr->u.tid_rdma.ack.kdeth1, JKEY, remote->jkey);
4887 ohdr->u.tid_rdma.ack.verbs_qp = cpu_to_be32(qp->remote_qpn);
4891 generation = kern_flow_generation_next(flow->flow_state.generation);
4892 *bth2 = mask_psn((generation << HFI1_KDETH_BTH_SEQ_SHIFT) - 1);
4893 qpriv->s_resync_psn = *bth2;
4894 *bth2 |= IB_BTH_REQ_ACK;
4895 KDETH_RESET(ohdr->u.tid_rdma.ack.kdeth0, KVER, 0x1);
4897 return sizeof(ohdr->u.tid_rdma.resync) / sizeof(u32);
4900 void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet)
4902 struct ib_other_headers *ohdr = packet->ohdr;
4903 struct rvt_qp *qp = packet->qp;
4904 struct hfi1_qp_priv *qpriv = qp->priv;
4905 struct hfi1_ctxtdata *rcd = qpriv->rcd;
4906 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
4907 struct rvt_ack_entry *e;
4908 struct tid_rdma_request *req;
4909 struct tid_rdma_flow *flow;
4910 struct tid_flow_state *fs = &qpriv->flow_state;
4911 u32 psn, generation, idx, gen_next;
4913 unsigned long flags;
4915 fecn = process_ecn(qp, packet);
4916 psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
4918 generation = mask_psn(psn + 1) >> HFI1_KDETH_BTH_SEQ_SHIFT;
4919 spin_lock_irqsave(&qp->s_lock, flags);
4921 gen_next = (fs->generation == KERN_GENERATION_RESERVED) ?
4922 generation : kern_flow_generation_next(fs->generation);
4924 * RESYNC packet contains the "next" generation and can only be
4925 * from the current or previous generations
4927 if (generation != mask_generation(gen_next - 1) &&
4928 generation != gen_next)
4930 /* Already processing a resync */
4934 spin_lock(&rcd->exp_lock);
4935 if (fs->index >= RXE_NUM_TID_FLOWS) {
4937 * If we don't have a flow, save the generation so it can be
4938 * applied when a new flow is allocated
4940 fs->generation = generation;
4942 /* Reprogram the QP flow with new generation */
4943 rcd->flows[fs->index].generation = generation;
4944 fs->generation = kern_setup_hw_flow(rcd, fs->index);
4948 * Disable SW PSN checking since a RESYNC is equivalent to a
4949 * sync point and the flow has/will be reprogrammed
4951 qpriv->s_flags &= ~HFI1_R_TID_SW_PSN;
4952 trace_hfi1_tid_write_rsp_rcv_resync(qp);
4955 * Reset all TID flow information with the new generation.
4956 * This is done for all requests and segments after the
4957 * last received segment
4959 for (idx = qpriv->r_tid_tail; ; idx++) {
4962 if (idx > rvt_size_atomic(&dev->rdi))
4964 e = &qp->s_ack_queue[idx];
4965 if (e->opcode == TID_OP(WRITE_REQ)) {
4966 req = ack_to_tid_req(e);
4967 trace_hfi1_tid_req_rcv_resync(qp, 0, e->opcode, e->psn,
4970 /* start from last unacked segment */
4971 for (flow_idx = req->clear_tail;
4972 CIRC_CNT(req->setup_head, flow_idx,
4974 flow_idx = CIRC_NEXT(flow_idx, MAX_FLOWS)) {
4978 flow = &req->flows[flow_idx];
4979 lpsn = full_flow_psn(flow,
4980 flow->flow_state.lpsn);
4981 next = flow->flow_state.r_next_psn;
4982 flow->npkts = delta_psn(lpsn, next - 1);
4983 flow->flow_state.generation = fs->generation;
4984 flow->flow_state.spsn = fs->psn;
4985 flow->flow_state.lpsn =
4986 flow->flow_state.spsn + flow->npkts - 1;
4987 flow->flow_state.r_next_psn =
4989 flow->flow_state.spsn);
4990 fs->psn += flow->npkts;
4991 trace_hfi1_tid_flow_rcv_resync(qp, flow_idx,
4995 if (idx == qp->s_tail_ack_queue)
4999 spin_unlock(&rcd->exp_lock);
5000 qpriv->resync = true;
5001 /* RESYNC request always gets a TID RDMA ACK. */
5002 qpriv->s_nak_state = 0;
5003 qpriv->s_flags |= RVT_S_ACK_PENDING;
5004 hfi1_schedule_tid_send(qp);
5007 qp->s_flags |= RVT_S_ECN;
5008 spin_unlock_irqrestore(&qp->s_lock, flags);
5012 * Call this function when the last TID RDMA WRITE DATA packet for a request
5015 static void update_tid_tail(struct rvt_qp *qp)
5016 __must_hold(&qp->s_lock)
5018 struct hfi1_qp_priv *priv = qp->priv;
5020 struct rvt_swqe *wqe;
5022 lockdep_assert_held(&qp->s_lock);
5023 /* Can't move beyond s_tid_cur */
5024 if (priv->s_tid_tail == priv->s_tid_cur)
5026 for (i = priv->s_tid_tail + 1; ; i++) {
5027 if (i == qp->s_size)
5030 if (i == priv->s_tid_cur)
5032 wqe = rvt_get_swqe_ptr(qp, i);
5033 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE)
5036 priv->s_tid_tail = i;
5037 priv->s_state = TID_OP(WRITE_RESP);
5040 int hfi1_make_tid_rdma_pkt(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
5041 __must_hold(&qp->s_lock)
5043 struct hfi1_qp_priv *priv = qp->priv;
5044 struct rvt_swqe *wqe;
5045 u32 bth1 = 0, bth2 = 0, hwords = 5, len, middle = 0;
5046 struct ib_other_headers *ohdr;
5047 struct rvt_sge_state *ss = &qp->s_sge;
5048 struct rvt_ack_entry *e = &qp->s_ack_queue[qp->s_tail_ack_queue];
5049 struct tid_rdma_request *req = ack_to_tid_req(e);
5051 u8 opcode = TID_OP(WRITE_DATA);
5053 lockdep_assert_held(&qp->s_lock);
5054 trace_hfi1_tid_write_sender_make_tid_pkt(qp, 0);
5056 * Prioritize the sending of the requests and responses over the
5057 * sending of the TID RDMA data packets.
5059 if (((atomic_read(&priv->n_tid_requests) < HFI1_TID_RDMA_WRITE_CNT) &&
5060 atomic_read(&priv->n_requests) &&
5061 !(qp->s_flags & (RVT_S_BUSY | RVT_S_WAIT_ACK |
5062 HFI1_S_ANY_WAIT_IO))) ||
5063 (e->opcode == TID_OP(WRITE_REQ) && req->cur_seg < req->alloc_seg &&
5064 !(qp->s_flags & (RVT_S_BUSY | HFI1_S_ANY_WAIT_IO)))) {
5065 struct iowait_work *iowork;
5067 iowork = iowait_get_ib_work(&priv->s_iowait);
5068 ps->s_txreq = get_waiting_verbs_txreq(iowork);
5069 if (ps->s_txreq || hfi1_make_rc_req(qp, ps)) {
5070 priv->s_flags |= HFI1_S_TID_BUSY_SET;
5075 ps->s_txreq = get_txreq(ps->dev, qp);
5079 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
5081 if ((priv->s_flags & RVT_S_ACK_PENDING) &&
5082 make_tid_rdma_ack(qp, ohdr, ps))
5086 * Bail out if we can't send data.
5087 * Be reminded that this check must been done after the call to
5088 * make_tid_rdma_ack() because the responding QP could be in
5089 * RTR state where it can send TID RDMA ACK, not TID RDMA WRITE DATA.
5091 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK))
5094 if (priv->s_flags & RVT_S_WAIT_ACK)
5097 /* Check whether there is anything to do. */
5098 if (priv->s_tid_tail == HFI1_QP_WQE_INVALID)
5100 wqe = rvt_get_swqe_ptr(qp, priv->s_tid_tail);
5101 req = wqe_to_tid_req(wqe);
5102 trace_hfi1_tid_req_make_tid_pkt(qp, 0, wqe->wr.opcode, wqe->psn,
5104 switch (priv->s_state) {
5105 case TID_OP(WRITE_REQ):
5106 case TID_OP(WRITE_RESP):
5107 priv->tid_ss.sge = wqe->sg_list[0];
5108 priv->tid_ss.sg_list = wqe->sg_list + 1;
5109 priv->tid_ss.num_sge = wqe->wr.num_sge;
5110 priv->tid_ss.total_len = wqe->length;
5112 if (priv->s_state == TID_OP(WRITE_REQ))
5113 hfi1_tid_rdma_restart_req(qp, wqe, &bth2);
5114 priv->s_state = TID_OP(WRITE_DATA);
5117 case TID_OP(WRITE_DATA):
5119 * 1. Check whether TID RDMA WRITE RESP available.
5121 * 2.1 If have more segments and no TID RDMA WRITE RESP,
5122 * set HFI1_S_WAIT_TID_RESP
5123 * 2.2 Return indicating no progress made.
5125 * 3.1 Build TID RDMA WRITE DATA packet.
5126 * 3.2 If last packet in segment:
5127 * 3.2.1 Change KDETH header bits
5128 * 3.2.2 Advance RESP pointers.
5129 * 3.3 Return indicating progress made.
5131 trace_hfi1_sender_make_tid_pkt(qp);
5132 trace_hfi1_tid_write_sender_make_tid_pkt(qp, 0);
5133 wqe = rvt_get_swqe_ptr(qp, priv->s_tid_tail);
5134 req = wqe_to_tid_req(wqe);
5137 if (!req->comp_seg || req->cur_seg == req->comp_seg)
5140 trace_hfi1_tid_req_make_tid_pkt(qp, 0, wqe->wr.opcode,
5141 wqe->psn, wqe->lpsn, req);
5142 last = hfi1_build_tid_rdma_packet(wqe, ohdr, &bth1, &bth2,
5146 /* move pointer to next flow */
5147 req->clear_tail = CIRC_NEXT(req->clear_tail,
5149 if (++req->cur_seg < req->total_segs) {
5150 if (!CIRC_CNT(req->setup_head, req->clear_tail,
5152 qp->s_flags |= HFI1_S_WAIT_TID_RESP;
5154 priv->s_state = TID_OP(WRITE_DATA_LAST);
5155 opcode = TID_OP(WRITE_DATA_LAST);
5157 /* Advance the s_tid_tail now */
5158 update_tid_tail(qp);
5161 hwords += sizeof(ohdr->u.tid_rdma.w_data) / sizeof(u32);
5165 case TID_OP(RESYNC):
5166 trace_hfi1_sender_make_tid_pkt(qp);
5167 /* Use generation from the most recently received response */
5168 wqe = rvt_get_swqe_ptr(qp, priv->s_tid_cur);
5169 req = wqe_to_tid_req(wqe);
5170 /* If no responses for this WQE look at the previous one */
5171 if (!req->comp_seg) {
5172 wqe = rvt_get_swqe_ptr(qp,
5173 (!priv->s_tid_cur ? qp->s_size :
5174 priv->s_tid_cur) - 1);
5175 req = wqe_to_tid_req(wqe);
5177 hwords += hfi1_build_tid_rdma_resync(qp, wqe, ohdr, &bth1,
5179 CIRC_PREV(req->setup_head,
5183 opcode = TID_OP(RESYNC);
5189 if (priv->s_flags & RVT_S_SEND_ONE) {
5190 priv->s_flags &= ~RVT_S_SEND_ONE;
5191 priv->s_flags |= RVT_S_WAIT_ACK;
5192 bth2 |= IB_BTH_REQ_ACK;
5195 ps->s_txreq->hdr_dwords = hwords;
5196 ps->s_txreq->sde = priv->s_sde;
5197 ps->s_txreq->ss = ss;
5198 ps->s_txreq->s_cur_size = len;
5199 hfi1_make_ruc_header(qp, ohdr, (opcode << 24), bth1, bth2,
5203 hfi1_put_txreq(ps->s_txreq);
5206 priv->s_flags &= ~RVT_S_BUSY;
5208 * If we didn't get a txreq, the QP will be woken up later to try
5209 * again, set the flags to the the wake up which work item to wake
5211 * (A better algorithm should be found to do this and generalize the
5212 * sleep/wakeup flags.)
5214 iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
5218 static int make_tid_rdma_ack(struct rvt_qp *qp,
5219 struct ib_other_headers *ohdr,
5220 struct hfi1_pkt_state *ps)
5222 struct rvt_ack_entry *e;
5223 struct hfi1_qp_priv *qpriv = qp->priv;
5224 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
5227 u32 bth1 = 0, bth2 = 0;
5230 struct tid_rdma_request *req, *nreq;
5232 trace_hfi1_tid_write_rsp_make_tid_ack(qp);
5233 /* Don't send an ACK if we aren't supposed to. */
5234 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
5237 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
5240 e = &qp->s_ack_queue[qpriv->r_tid_ack];
5241 req = ack_to_tid_req(e);
5243 * In the RESYNC case, we are exactly one segment past the
5244 * previously sent ack or at the previously sent NAK. So to send
5245 * the resync ack, we go back one segment (which might be part of
5246 * the previous request) and let the do-while loop execute again.
5247 * The advantage of executing the do-while loop is that any data
5248 * received after the previous ack is automatically acked in the
5249 * RESYNC ack. It turns out that for the do-while loop we only need
5250 * to pull back qpriv->r_tid_ack, not the segment
5251 * indices/counters. The scheme works even if the previous request
5252 * was not a TID WRITE request.
5254 if (qpriv->resync) {
5255 if (!req->ack_seg || req->ack_seg == req->total_segs)
5256 qpriv->r_tid_ack = !qpriv->r_tid_ack ?
5257 rvt_size_atomic(&dev->rdi) :
5258 qpriv->r_tid_ack - 1;
5259 e = &qp->s_ack_queue[qpriv->r_tid_ack];
5260 req = ack_to_tid_req(e);
5263 trace_hfi1_rsp_make_tid_ack(qp, e->psn);
5264 trace_hfi1_tid_req_make_tid_ack(qp, 0, e->opcode, e->psn, e->lpsn,
5267 * If we've sent all the ACKs that we can, we are done
5268 * until we get more segments...
5270 if (!qpriv->s_nak_state && !qpriv->resync &&
5271 req->ack_seg == req->comp_seg)
5276 * To deal with coalesced ACKs, the acked_tail pointer
5277 * into the flow array is used. The distance between it
5278 * and the clear_tail is the number of flows that are
5282 /* Get up-to-date value */
5283 CIRC_CNT(req->clear_tail, req->acked_tail,
5285 /* Advance acked index */
5286 req->acked_tail = req->clear_tail;
5289 * req->clear_tail points to the segment currently being
5290 * received. So, when sending an ACK, the previous
5291 * segment is being ACK'ed.
5293 flow = CIRC_PREV(req->acked_tail, MAX_FLOWS);
5294 if (req->ack_seg != req->total_segs)
5296 req->state = TID_REQUEST_COMPLETE;
5298 next = qpriv->r_tid_ack + 1;
5299 if (next > rvt_size_atomic(&dev->rdi))
5301 qpriv->r_tid_ack = next;
5302 if (qp->s_ack_queue[next].opcode != TID_OP(WRITE_REQ))
5304 nreq = ack_to_tid_req(&qp->s_ack_queue[next]);
5305 if (!nreq->comp_seg || nreq->ack_seg == nreq->comp_seg)
5308 /* Move to the next ack entry now */
5309 e = &qp->s_ack_queue[qpriv->r_tid_ack];
5310 req = ack_to_tid_req(e);
5314 * At this point qpriv->r_tid_ack == qpriv->r_tid_tail but e and
5315 * req could be pointing at the previous ack queue entry
5317 if (qpriv->s_nak_state ||
5319 !hfi1_tid_rdma_is_resync_psn(qpriv->r_next_psn_kdeth - 1) &&
5320 (cmp_psn(qpriv->r_next_psn_kdeth - 1,
5321 full_flow_psn(&req->flows[flow],
5322 req->flows[flow].flow_state.lpsn)) > 0))) {
5324 * A NAK will implicitly acknowledge all previous TID RDMA
5325 * requests. Therefore, we NAK with the req->acked_tail
5326 * segment for the request at qpriv->r_tid_ack (same at
5327 * this point as the req->clear_tail segment for the
5328 * qpriv->r_tid_tail request)
5330 e = &qp->s_ack_queue[qpriv->r_tid_ack];
5331 req = ack_to_tid_req(e);
5332 flow = req->acked_tail;
5333 } else if (req->ack_seg == req->total_segs &&
5334 qpriv->s_flags & HFI1_R_TID_WAIT_INTERLCK)
5335 qpriv->s_flags &= ~HFI1_R_TID_WAIT_INTERLCK;
5337 trace_hfi1_tid_write_rsp_make_tid_ack(qp);
5338 trace_hfi1_tid_req_make_tid_ack(qp, 0, e->opcode, e->psn, e->lpsn,
5340 hwords += hfi1_build_tid_rdma_write_ack(qp, e, ohdr, flow, &bth1,
5343 qpriv->s_flags &= ~RVT_S_ACK_PENDING;
5344 ps->s_txreq->hdr_dwords = hwords;
5345 ps->s_txreq->sde = qpriv->s_sde;
5346 ps->s_txreq->s_cur_size = len;
5347 ps->s_txreq->ss = NULL;
5348 hfi1_make_ruc_header(qp, ohdr, (TID_OP(ACK) << 24), bth1, bth2, middle,
5350 ps->s_txreq->txreq.flags |= SDMA_TXREQ_F_VIP;
5354 * Ensure s_rdma_ack_cnt changes are committed prior to resetting
5355 * RVT_S_RESP_PENDING
5358 qpriv->s_flags &= ~RVT_S_ACK_PENDING;
5362 static int hfi1_send_tid_ok(struct rvt_qp *qp)
5364 struct hfi1_qp_priv *priv = qp->priv;
5366 return !(priv->s_flags & RVT_S_BUSY ||
5367 qp->s_flags & HFI1_S_ANY_WAIT_IO) &&
5368 (verbs_txreq_queued(iowait_get_tid_work(&priv->s_iowait)) ||
5369 (priv->s_flags & RVT_S_RESP_PENDING) ||
5370 !(qp->s_flags & HFI1_S_ANY_TID_WAIT_SEND));
5373 void _hfi1_do_tid_send(struct work_struct *work)
5375 struct iowait_work *w = container_of(work, struct iowait_work, iowork);
5376 struct rvt_qp *qp = iowait_to_qp(w->iow);
5378 hfi1_do_tid_send(qp);
5381 static void hfi1_do_tid_send(struct rvt_qp *qp)
5383 struct hfi1_pkt_state ps;
5384 struct hfi1_qp_priv *priv = qp->priv;
5386 ps.dev = to_idev(qp->ibqp.device);
5387 ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
5388 ps.ppd = ppd_from_ibp(ps.ibp);
5389 ps.wait = iowait_get_tid_work(&priv->s_iowait);
5390 ps.in_thread = false;
5391 ps.timeout_int = qp->timeout_jiffies / 8;
5393 trace_hfi1_rc_do_tid_send(qp, false);
5394 spin_lock_irqsave(&qp->s_lock, ps.flags);
5396 /* Return if we are already busy processing a work request. */
5397 if (!hfi1_send_tid_ok(qp)) {
5398 if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
5399 iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
5400 spin_unlock_irqrestore(&qp->s_lock, ps.flags);
5404 priv->s_flags |= RVT_S_BUSY;
5406 ps.timeout = jiffies + ps.timeout_int;
5407 ps.cpu = priv->s_sde ? priv->s_sde->cpu :
5408 cpumask_first(cpumask_of_node(ps.ppd->dd->node));
5409 ps.pkts_sent = false;
5411 /* insure a pre-built packet is handled */
5412 ps.s_txreq = get_waiting_verbs_txreq(ps.wait);
5414 /* Check for a constructed packet to be sent. */
5416 if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
5417 qp->s_flags |= RVT_S_BUSY;
5418 ps.wait = iowait_get_ib_work(&priv->s_iowait);
5420 spin_unlock_irqrestore(&qp->s_lock, ps.flags);
5423 * If the packet cannot be sent now, return and
5424 * the send tasklet will be woken up later.
5426 if (hfi1_verbs_send(qp, &ps))
5429 /* allow other tasks to run */
5430 if (hfi1_schedule_send_yield(qp, &ps, true))
5433 spin_lock_irqsave(&qp->s_lock, ps.flags);
5434 if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
5435 qp->s_flags &= ~RVT_S_BUSY;
5436 priv->s_flags &= ~HFI1_S_TID_BUSY_SET;
5437 ps.wait = iowait_get_tid_work(&priv->s_iowait);
5438 if (iowait_flag_set(&priv->s_iowait,
5440 hfi1_schedule_send(qp);
5443 } while (hfi1_make_tid_rdma_pkt(qp, &ps));
5444 iowait_starve_clear(ps.pkts_sent, &priv->s_iowait);
5445 spin_unlock_irqrestore(&qp->s_lock, ps.flags);
5448 static bool _hfi1_schedule_tid_send(struct rvt_qp *qp)
5450 struct hfi1_qp_priv *priv = qp->priv;
5451 struct hfi1_ibport *ibp =
5452 to_iport(qp->ibqp.device, qp->port_num);
5453 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
5454 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
5456 return iowait_tid_schedule(&priv->s_iowait, ppd->hfi1_wq,
5459 cpumask_first(cpumask_of_node(dd->node)));
5463 * hfi1_schedule_tid_send - schedule progress on TID RDMA state machine
5466 * This schedules qp progress on the TID RDMA state machine. Caller
5467 * should hold the s_lock.
5468 * Unlike hfi1_schedule_send(), this cannot use hfi1_send_ok() because
5469 * the two state machines can step on each other with respect to the
5471 * Therefore, a modified test is used.
5472 * @return true if the second leg is scheduled;
5473 * false if the second leg is not scheduled.
5475 bool hfi1_schedule_tid_send(struct rvt_qp *qp)
5477 lockdep_assert_held(&qp->s_lock);
5478 if (hfi1_send_tid_ok(qp)) {
5480 * The following call returns true if the qp is not on the
5481 * queue and false if the qp is already on the queue before
5482 * this call. Either way, the qp will be on the queue when the
5485 _hfi1_schedule_tid_send(qp);
5488 if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
5489 iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
5490 IOWAIT_PENDING_TID);
5494 bool hfi1_tid_rdma_ack_interlock(struct rvt_qp *qp, struct rvt_ack_entry *e)
5496 struct rvt_ack_entry *prev;
5497 struct tid_rdma_request *req;
5498 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
5499 struct hfi1_qp_priv *priv = qp->priv;
5502 s_prev = qp->s_tail_ack_queue == 0 ? rvt_size_atomic(&dev->rdi) :
5503 (qp->s_tail_ack_queue - 1);
5504 prev = &qp->s_ack_queue[s_prev];
5506 if ((e->opcode == TID_OP(READ_REQ) ||
5507 e->opcode == OP(RDMA_READ_REQUEST)) &&
5508 prev->opcode == TID_OP(WRITE_REQ)) {
5509 req = ack_to_tid_req(prev);
5510 if (req->ack_seg != req->total_segs) {
5511 priv->s_flags |= HFI1_R_TID_WAIT_INTERLCK;
5518 static u32 read_r_next_psn(struct hfi1_devdata *dd, u8 ctxt, u8 fidx)
5523 * The only sane way to get the amount of
5524 * progress is to read the HW flow state.
5526 reg = read_uctxt_csr(dd, ctxt, RCV_TID_FLOW_TABLE + (8 * fidx));
5527 return mask_psn(reg);
5530 static void tid_rdma_rcv_err(struct hfi1_packet *packet,
5531 struct ib_other_headers *ohdr,
5532 struct rvt_qp *qp, u32 psn, int diff, bool fecn)
5534 unsigned long flags;
5536 tid_rdma_rcv_error(packet, ohdr, qp, psn, diff);
5538 spin_lock_irqsave(&qp->s_lock, flags);
5539 qp->s_flags |= RVT_S_ECN;
5540 spin_unlock_irqrestore(&qp->s_lock, flags);
5544 static void update_r_next_psn_fecn(struct hfi1_packet *packet,
5545 struct hfi1_qp_priv *priv,
5546 struct hfi1_ctxtdata *rcd,
5547 struct tid_rdma_flow *flow,
5551 * If a start/middle packet is delivered here due to
5552 * RSM rule and FECN, we need to update the r_next_psn.
5554 if (fecn && packet->etype == RHF_RCV_TYPE_EAGER &&
5555 !(priv->s_flags & HFI1_R_TID_SW_PSN)) {
5556 struct hfi1_devdata *dd = rcd->dd;
5558 flow->flow_state.r_next_psn =
5559 read_r_next_psn(dd, rcd->ctxt, flow->idx);