2 * IBM eServer eHCA Infiniband device driver for Linux on POWER
4 * post_send/recv, poll_cq, req_notify
6 * Authors: Waleri Fomin <fomin@de.ibm.com>
7 * Hoang-Nam Nguyen <hnguyen@de.ibm.com>
8 * Reinhard Ernst <rernst@de.ibm.com>
10 * Copyright (c) 2005 IBM Corporation
12 * All rights reserved.
14 * This source code is distributed under a dual license of GPL v2.0 and OpenIB
19 * Redistribution and use in source and binary forms, with or without
20 * modification, are permitted provided that the following conditions are met:
22 * Redistributions of source code must retain the above copyright notice, this
23 * list of conditions and the following disclaimer.
25 * Redistributions in binary form must reproduce the above copyright notice,
26 * this list of conditions and the following disclaimer in the documentation
27 * and/or other materials
28 * provided with the distribution.
30 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
31 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
34 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
35 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
36 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
37 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER
38 * IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
39 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
40 * POSSIBILITY OF SUCH DAMAGE.
44 #include <asm-powerpc/system.h>
45 #include "ehca_classes.h"
46 #include "ehca_tools.h"
48 #include "ehca_iverbs.h"
52 static inline int ehca_write_rwqe(struct ipz_queue *ipz_rqueue,
53 struct ehca_wqe *wqe_p,
54 struct ib_recv_wr *recv_wr)
57 if (unlikely((recv_wr->num_sge < 0) ||
58 (recv_wr->num_sge > ipz_rqueue->act_nr_of_sg))) {
59 ehca_gen_err("Invalid number of WQE SGE. "
60 "num_sqe=%x max_nr_of_sg=%x",
61 recv_wr->num_sge, ipz_rqueue->act_nr_of_sg);
62 return -EINVAL; /* invalid SG list length */
65 /* clear wqe header until sglist */
66 memset(wqe_p, 0, offsetof(struct ehca_wqe, u.ud_av.sg_list));
68 wqe_p->work_request_id = recv_wr->wr_id;
69 wqe_p->nr_of_data_seg = recv_wr->num_sge;
71 for (cnt_ds = 0; cnt_ds < recv_wr->num_sge; cnt_ds++) {
72 wqe_p->u.all_rcv.sg_list[cnt_ds].vaddr =
73 recv_wr->sg_list[cnt_ds].addr;
74 wqe_p->u.all_rcv.sg_list[cnt_ds].lkey =
75 recv_wr->sg_list[cnt_ds].lkey;
76 wqe_p->u.all_rcv.sg_list[cnt_ds].length =
77 recv_wr->sg_list[cnt_ds].length;
80 if (ehca_debug_level) {
81 ehca_gen_dbg("RECEIVE WQE written into ipz_rqueue=%p", ipz_rqueue);
82 ehca_dmp( wqe_p, 16*(6 + wqe_p->nr_of_data_seg), "recv wqe");
88 #if defined(DEBUG_GSI_SEND_WR)
90 /* need ib_mad struct */
91 #include <rdma/ib_mad.h>
93 static void trace_send_wr_ud(const struct ib_send_wr *send_wr)
98 struct ib_mad_hdr *mad_hdr = send_wr->wr.ud.mad_hdr;
99 struct ib_sge *sge = send_wr->sg_list;
100 ehca_gen_dbg("send_wr#%x wr_id=%lx num_sge=%x "
101 "send_flags=%x opcode=%x",idx, send_wr->wr_id,
102 send_wr->num_sge, send_wr->send_flags,
105 ehca_gen_dbg("send_wr#%x mad_hdr base_version=%x "
106 "mgmt_class=%x class_version=%x method=%x "
107 "status=%x class_specific=%x tid=%lx "
108 "attr_id=%x resv=%x attr_mod=%x",
109 idx, mad_hdr->base_version,
111 mad_hdr->class_version, mad_hdr->method,
112 mad_hdr->status, mad_hdr->class_specific,
113 mad_hdr->tid, mad_hdr->attr_id,
117 for (j = 0; j < send_wr->num_sge; j++) {
118 u8 *data = (u8 *) abs_to_virt(sge->addr);
119 ehca_gen_dbg("send_wr#%x sge#%x addr=%p length=%x "
121 idx, j, data, sge->length, sge->lkey);
122 /* assume length is n*16 */
123 ehca_dmp(data, sge->length, "send_wr#%x sge#%x",
128 send_wr = send_wr->next;
129 } /* eof while send_wr */
132 #endif /* DEBUG_GSI_SEND_WR */
134 static inline int ehca_write_swqe(struct ehca_qp *qp,
135 struct ehca_wqe *wqe_p,
136 const struct ib_send_wr *send_wr)
140 struct ehca_av *my_av;
141 u32 remote_qkey = send_wr->wr.ud.remote_qkey;
143 if (unlikely((send_wr->num_sge < 0) ||
144 (send_wr->num_sge > qp->ipz_squeue.act_nr_of_sg))) {
145 ehca_gen_err("Invalid number of WQE SGE. "
146 "num_sqe=%x max_nr_of_sg=%x",
147 send_wr->num_sge, qp->ipz_squeue.act_nr_of_sg);
148 return -EINVAL; /* invalid SG list length */
151 /* clear wqe header until sglist */
152 memset(wqe_p, 0, offsetof(struct ehca_wqe, u.ud_av.sg_list));
154 wqe_p->work_request_id = send_wr->wr_id;
156 switch (send_wr->opcode) {
158 case IB_WR_SEND_WITH_IMM:
159 wqe_p->optype = WQE_OPTYPE_SEND;
161 case IB_WR_RDMA_WRITE:
162 case IB_WR_RDMA_WRITE_WITH_IMM:
163 wqe_p->optype = WQE_OPTYPE_RDMAWRITE;
165 case IB_WR_RDMA_READ:
166 wqe_p->optype = WQE_OPTYPE_RDMAREAD;
169 ehca_gen_err("Invalid opcode=%x", send_wr->opcode);
170 return -EINVAL; /* invalid opcode */
173 wqe_p->wqef = (send_wr->opcode) & WQEF_HIGH_NIBBLE;
177 if (send_wr->send_flags & IB_SEND_SIGNALED)
178 wqe_p->wr_flag |= WQE_WRFLAG_REQ_SIGNAL_COM;
180 if (send_wr->opcode == IB_WR_SEND_WITH_IMM ||
181 send_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
182 /* this might not work as long as HW does not support it */
183 wqe_p->immediate_data = be32_to_cpu(send_wr->imm_data);
184 wqe_p->wr_flag |= WQE_WRFLAG_IMM_DATA_PRESENT;
187 wqe_p->nr_of_data_seg = send_wr->num_sge;
189 switch (qp->qp_type) {
192 /* no break is intential here */
194 /* IB 1.2 spec C10-15 compliance */
195 if (send_wr->wr.ud.remote_qkey & 0x80000000)
196 remote_qkey = qp->qkey;
198 wqe_p->destination_qp_number = send_wr->wr.ud.remote_qpn << 8;
199 wqe_p->local_ee_context_qkey = remote_qkey;
200 if (!send_wr->wr.ud.ah) {
201 ehca_gen_err("wr.ud.ah is NULL. qp=%p", qp);
204 my_av = container_of(send_wr->wr.ud.ah, struct ehca_av, ib_ah);
205 wqe_p->u.ud_av.ud_av = my_av->av;
208 * omitted check of IB_SEND_INLINE
209 * since HW does not support it
211 for (idx = 0; idx < send_wr->num_sge; idx++) {
212 wqe_p->u.ud_av.sg_list[idx].vaddr =
213 send_wr->sg_list[idx].addr;
214 wqe_p->u.ud_av.sg_list[idx].lkey =
215 send_wr->sg_list[idx].lkey;
216 wqe_p->u.ud_av.sg_list[idx].length =
217 send_wr->sg_list[idx].length;
219 if (qp->qp_type == IB_QPT_SMI ||
220 qp->qp_type == IB_QPT_GSI)
221 wqe_p->u.ud_av.ud_av.pmtu = 1;
222 if (qp->qp_type == IB_QPT_GSI) {
223 wqe_p->pkeyi = send_wr->wr.ud.pkey_index;
224 #ifdef DEBUG_GSI_SEND_WR
225 trace_send_wr_ud(send_wr);
226 #endif /* DEBUG_GSI_SEND_WR */
231 if (send_wr->send_flags & IB_SEND_FENCE)
232 wqe_p->wr_flag |= WQE_WRFLAG_FENCE;
233 /* no break is intentional here */
235 /* TODO: atomic not implemented */
236 wqe_p->u.nud.remote_virtual_adress =
237 send_wr->wr.rdma.remote_addr;
238 wqe_p->u.nud.rkey = send_wr->wr.rdma.rkey;
241 * omitted checking of IB_SEND_INLINE
242 * since HW does not support it
245 for (idx = 0; idx < send_wr->num_sge; idx++) {
246 wqe_p->u.nud.sg_list[idx].vaddr =
247 send_wr->sg_list[idx].addr;
248 wqe_p->u.nud.sg_list[idx].lkey =
249 send_wr->sg_list[idx].lkey;
250 wqe_p->u.nud.sg_list[idx].length =
251 send_wr->sg_list[idx].length;
252 dma_length += send_wr->sg_list[idx].length;
254 wqe_p->u.nud.atomic_1st_op_dma_len = dma_length;
259 ehca_gen_err("Invalid qptype=%x", qp->qp_type);
263 if (ehca_debug_level) {
264 ehca_gen_dbg("SEND WQE written into queue qp=%p ", qp);
265 ehca_dmp( wqe_p, 16*(6 + wqe_p->nr_of_data_seg), "send wqe");
270 /* map_ib_wc_status converts raw cqe_status to ib_wc_status */
271 static inline void map_ib_wc_status(u32 cqe_status,
272 enum ib_wc_status *wc_status)
274 if (unlikely(cqe_status & WC_STATUS_ERROR_BIT)) {
275 switch (cqe_status & 0x3F) {
278 *wc_status = IB_WC_LOC_LEN_ERR;
282 *wc_status = IB_WC_LOC_QP_OP_ERR;
286 *wc_status = IB_WC_LOC_EEC_OP_ERR;
290 *wc_status = IB_WC_LOC_PROT_ERR;
294 *wc_status = IB_WC_WR_FLUSH_ERR;
297 *wc_status = IB_WC_MW_BIND_ERR;
299 case 0x07: /* remote error - look into bits 20:24 */
301 & WC_STATUS_REMOTE_ERROR_FLAGS) >> 11) {
304 * PSN Sequence Error!
305 * couldn't find a matching status!
307 *wc_status = IB_WC_GENERAL_ERR;
310 *wc_status = IB_WC_REM_INV_REQ_ERR;
313 *wc_status = IB_WC_REM_ACCESS_ERR;
316 *wc_status = IB_WC_REM_OP_ERR;
319 *wc_status = IB_WC_REM_INV_RD_REQ_ERR;
324 *wc_status = IB_WC_RETRY_EXC_ERR;
327 *wc_status = IB_WC_RNR_RETRY_EXC_ERR;
331 *wc_status = IB_WC_REM_ABORT_ERR;
335 *wc_status = IB_WC_INV_EECN_ERR;
339 *wc_status = IB_WC_INV_EEC_STATE_ERR;
342 *wc_status = IB_WC_BAD_RESP_ERR;
346 *wc_status = IB_WC_WR_FLUSH_ERR;
349 *wc_status = IB_WC_FATAL_ERR;
353 *wc_status = IB_WC_SUCCESS;
356 int ehca_post_send(struct ib_qp *qp,
357 struct ib_send_wr *send_wr,
358 struct ib_send_wr **bad_send_wr)
360 struct ehca_qp *my_qp = container_of(qp, struct ehca_qp, ib_qp);
361 struct ib_send_wr *cur_send_wr;
362 struct ehca_wqe *wqe_p;
365 unsigned long spl_flags;
368 spin_lock_irqsave(&my_qp->spinlock_s, spl_flags);
370 /* loop processes list of send reqs */
371 for (cur_send_wr = send_wr; cur_send_wr != NULL;
372 cur_send_wr = cur_send_wr->next) {
373 u64 start_offset = my_qp->ipz_squeue.current_q_offset;
374 /* get pointer next to free WQE */
375 wqe_p = ipz_qeit_get_inc(&my_qp->ipz_squeue);
376 if (unlikely(!wqe_p)) {
377 /* too many posted work requests: queue overflow */
379 *bad_send_wr = cur_send_wr;
382 ehca_err(qp->device, "Too many posted WQEs "
383 "qp_num=%x", qp->qp_num);
385 goto post_send_exit0;
387 /* write a SEND WQE into the QUEUE */
388 ret = ehca_write_swqe(my_qp, wqe_p, cur_send_wr);
390 * if something failed,
391 * reset the free entry pointer to the start value
394 my_qp->ipz_squeue.current_q_offset = start_offset;
395 *bad_send_wr = cur_send_wr;
398 ehca_err(qp->device, "Could not write WQE "
399 "qp_num=%x", qp->qp_num);
401 goto post_send_exit0;
404 ehca_dbg(qp->device, "ehca_qp=%p qp_num=%x wqe_cnt=%d",
405 my_qp, qp->qp_num, wqe_cnt);
406 } /* eof for cur_send_wr */
409 /* UNLOCK the QUEUE */
410 spin_unlock_irqrestore(&my_qp->spinlock_s, spl_flags);
411 iosync(); /* serialize GAL register access */
412 hipz_update_sqa(my_qp, wqe_cnt);
416 int ehca_post_recv(struct ib_qp *qp,
417 struct ib_recv_wr *recv_wr,
418 struct ib_recv_wr **bad_recv_wr)
420 struct ehca_qp *my_qp = container_of(qp, struct ehca_qp, ib_qp);
421 struct ib_recv_wr *cur_recv_wr;
422 struct ehca_wqe *wqe_p;
425 unsigned long spl_flags;
428 spin_lock_irqsave(&my_qp->spinlock_r, spl_flags);
430 /* loop processes list of send reqs */
431 for (cur_recv_wr = recv_wr; cur_recv_wr != NULL;
432 cur_recv_wr = cur_recv_wr->next) {
433 u64 start_offset = my_qp->ipz_rqueue.current_q_offset;
434 /* get pointer next to free WQE */
435 wqe_p = ipz_qeit_get_inc(&my_qp->ipz_rqueue);
436 if (unlikely(!wqe_p)) {
437 /* too many posted work requests: queue overflow */
439 *bad_recv_wr = cur_recv_wr;
442 ehca_err(qp->device, "Too many posted WQEs "
443 "qp_num=%x", qp->qp_num);
445 goto post_recv_exit0;
447 /* write a RECV WQE into the QUEUE */
448 ret = ehca_write_rwqe(&my_qp->ipz_rqueue, wqe_p, cur_recv_wr);
450 * if something failed,
451 * reset the free entry pointer to the start value
454 my_qp->ipz_rqueue.current_q_offset = start_offset;
455 *bad_recv_wr = cur_recv_wr;
458 ehca_err(qp->device, "Could not write WQE "
459 "qp_num=%x", qp->qp_num);
461 goto post_recv_exit0;
464 ehca_gen_dbg("ehca_qp=%p qp_num=%x wqe_cnt=%d",
465 my_qp, qp->qp_num, wqe_cnt);
466 } /* eof for cur_recv_wr */
469 spin_unlock_irqrestore(&my_qp->spinlock_r, spl_flags);
470 iosync(); /* serialize GAL register access */
471 hipz_update_rqa(my_qp, wqe_cnt);
476 * ib_wc_opcode table converts ehca wc opcode to ib
477 * Since we use zero to indicate invalid opcode, the actual ib opcode must
480 static const u8 ib_wc_opcode[255] = {
481 [0x01] = IB_WC_RECV+1,
482 [0x02] = IB_WC_RECV_RDMA_WITH_IMM+1,
483 [0x04] = IB_WC_BIND_MW+1,
484 [0x08] = IB_WC_FETCH_ADD+1,
485 [0x10] = IB_WC_COMP_SWAP+1,
486 [0x20] = IB_WC_RDMA_WRITE+1,
487 [0x40] = IB_WC_RDMA_READ+1,
488 [0x80] = IB_WC_SEND+1
491 /* internal function to poll one entry of cq */
492 static inline int ehca_poll_cq_one(struct ib_cq *cq, struct ib_wc *wc)
495 struct ehca_cq *my_cq = container_of(cq, struct ehca_cq, ib_cq);
496 struct ehca_cqe *cqe;
499 poll_cq_one_read_cqe:
500 cqe = (struct ehca_cqe *)
501 ipz_qeit_get_inc_valid(&my_cq->ipz_queue);
504 ehca_dbg(cq->device, "Completion queue is empty ehca_cq=%p "
505 "cq_num=%x ret=%x", my_cq, my_cq->cq_number, ret);
506 goto poll_cq_one_exit0;
509 /* prevents loads being reordered across this point */
513 if (unlikely(cqe->status & WC_STATUS_PURGE_BIT)) {
514 struct ehca_qp *qp=ehca_cq_get_qp(my_cq, cqe->local_qp_number);
516 unsigned long spl_flags;
518 ehca_err(cq->device, "cq_num=%x qp_num=%x "
519 "could not find qp -> ignore cqe",
520 my_cq->cq_number, cqe->local_qp_number);
521 ehca_dmp(cqe, 64, "cq_num=%x qp_num=%x",
522 my_cq->cq_number, cqe->local_qp_number);
523 /* ignore this purged cqe */
524 goto poll_cq_one_read_cqe;
526 spin_lock_irqsave(&qp->spinlock_s, spl_flags);
527 purgeflag = qp->sqerr_purgeflag;
528 spin_unlock_irqrestore(&qp->spinlock_s, spl_flags);
531 ehca_dbg(cq->device, "Got CQE with purged bit qp_num=%x "
533 cqe->local_qp_number, cqe->remote_qp_number);
534 if (ehca_debug_level)
535 ehca_dmp(cqe, 64, "qp_num=%x src_qp=%x",
536 cqe->local_qp_number,
537 cqe->remote_qp_number);
539 * ignore this to avoid double cqes of bad wqe
540 * that caused sqe and turn off purge flag
542 qp->sqerr_purgeflag = 0;
543 goto poll_cq_one_read_cqe;
548 if (ehca_debug_level) {
550 "Received COMPLETION ehca_cq=%p cq_num=%x -----",
551 my_cq, my_cq->cq_number);
552 ehca_dmp(cqe, 64, "ehca_cq=%p cq_num=%x",
553 my_cq, my_cq->cq_number);
555 "ehca_cq=%p cq_num=%x -------------------------",
556 my_cq, my_cq->cq_number);
559 /* we got a completion! */
560 wc->wr_id = cqe->work_request_id;
562 /* eval ib_wc_opcode */
563 wc->opcode = ib_wc_opcode[cqe->optype]-1;
564 if (unlikely(wc->opcode == -1)) {
565 ehca_err(cq->device, "Invalid cqe->OPType=%x cqe->status=%x "
566 "ehca_cq=%p cq_num=%x",
567 cqe->optype, cqe->status, my_cq, my_cq->cq_number);
568 /* dump cqe for other infos */
569 ehca_dmp(cqe, 64, "ehca_cq=%p cq_num=%x",
570 my_cq, my_cq->cq_number);
571 /* update also queue adder to throw away this entry!!! */
572 goto poll_cq_one_exit0;
574 /* eval ib_wc_status */
575 if (unlikely(cqe->status & WC_STATUS_ERROR_BIT)) {
576 /* complete with errors */
577 map_ib_wc_status(cqe->status, &wc->status);
578 wc->vendor_err = wc->status;
580 wc->status = IB_WC_SUCCESS;
583 wc->byte_len = cqe->nr_bytes_transferred;
584 wc->pkey_index = cqe->pkey_index;
585 wc->slid = cqe->rlid;
586 wc->dlid_path_bits = cqe->dlid;
587 wc->src_qp = cqe->remote_qp_number;
588 wc->wc_flags = cqe->w_completion_flags;
589 wc->imm_data = cpu_to_be32(cqe->immediate_data);
590 wc->sl = cqe->service_level;
592 if (wc->status != IB_WC_SUCCESS)
594 "ehca_cq=%p cq_num=%x WARNING unsuccessful cqe "
595 "OPType=%x status=%x qp_num=%x src_qp=%x wr_id=%lx "
596 "cqe=%p", my_cq, my_cq->cq_number, cqe->optype,
597 cqe->status, cqe->local_qp_number,
598 cqe->remote_qp_number, cqe->work_request_id, cqe);
602 hipz_update_feca(my_cq, cqe_count);
607 int ehca_poll_cq(struct ib_cq *cq, int num_entries, struct ib_wc *wc)
609 struct ehca_cq *my_cq = container_of(cq, struct ehca_cq, ib_cq);
611 struct ib_wc *current_wc = wc;
613 unsigned long spl_flags;
615 if (num_entries < 1) {
616 ehca_err(cq->device, "Invalid num_entries=%d ehca_cq=%p "
617 "cq_num=%x", num_entries, my_cq, my_cq->cq_number);
622 spin_lock_irqsave(&my_cq->spinlock, spl_flags);
623 for (nr = 0; nr < num_entries; nr++) {
624 ret = ehca_poll_cq_one(cq, current_wc);
629 spin_unlock_irqrestore(&my_cq->spinlock, spl_flags);
630 if (ret == -EAGAIN || !ret)
637 int ehca_req_notify_cq(struct ib_cq *cq, enum ib_cq_notify_flags notify_flags)
639 struct ehca_cq *my_cq = container_of(cq, struct ehca_cq, ib_cq);
640 unsigned long spl_flags;
643 switch (notify_flags & IB_CQ_SOLICITED_MASK) {
644 case IB_CQ_SOLICITED:
645 hipz_set_cqx_n0(my_cq, 1);
647 case IB_CQ_NEXT_COMP:
648 hipz_set_cqx_n1(my_cq, 1);
654 if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
655 spin_lock_irqsave(&my_cq->spinlock, spl_flags);
656 ret = ipz_qeit_is_valid(&my_cq->ipz_queue);
657 spin_unlock_irqrestore(&my_cq->spinlock, spl_flags);