2 * Copyright (c) 2009-2013 Chelsio, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
38 #include <sys/types.h>
39 #include <sys/malloc.h>
40 #include <sys/socket.h>
41 #include <sys/socketvar.h>
42 #include <sys/sockio.h>
43 #include <sys/taskqueue.h>
44 #include <netinet/in.h>
45 #include <net/route.h>
47 #include <netinet/in_systm.h>
48 #include <netinet/in_pcb.h>
49 #include <netinet/ip.h>
50 #include <netinet/ip_var.h>
51 #include <netinet/tcp_var.h>
52 #include <netinet/tcp.h>
53 #include <netinet/tcpip.h>
55 #include <netinet/toecore.h>
59 #include <linux/types.h>
61 #include "tom/t4_tom.h"
66 extern int db_delay_usecs;
67 extern int db_fc_threshold;
68 static void creds(struct toepcb *toep, size_t wrsize);
71 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
74 spin_lock_irqsave(&qhp->lock, flag);
75 qhp->attr.state = state;
76 spin_unlock_irqrestore(&qhp->lock, flag);
79 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
82 contigfree(sq->queue, sq->memsize, M_DEVBUF);
85 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
88 dealloc_host_sq(rdev, sq);
91 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
93 sq->queue = contigmalloc(sq->memsize, M_DEVBUF, M_NOWAIT, 0ul, ~0ul,
97 sq->dma_addr = vtophys(sq->queue);
100 sq->phys_addr = vtophys(sq->queue);
101 pci_unmap_addr_set(sq, mapping, sq->dma_addr);
102 CTR4(KTR_IW_CXGBE, "%s sq %p dma_addr %p phys_addr %p", __func__,
103 sq->queue, sq->dma_addr, sq->phys_addr);
107 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
108 struct c4iw_dev_ucontext *uctx)
111 * uP clears EQ contexts when the connection exits rdma mode,
112 * so no need to post a RESET WR for these EQs.
114 contigfree(wq->rq.queue, wq->rq.memsize, M_DEVBUF);
115 dealloc_sq(rdev, &wq->sq);
116 c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
119 c4iw_put_qpid(rdev, wq->rq.qid, uctx);
120 c4iw_put_qpid(rdev, wq->sq.qid, uctx);
124 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
125 struct t4_cq *rcq, struct t4_cq *scq,
126 struct c4iw_dev_ucontext *uctx)
128 struct adapter *sc = rdev->adap;
129 int user = (uctx != &rdev->uctx);
130 struct fw_ri_res_wr *res_wr;
131 struct fw_ri_res *res;
133 struct c4iw_wr_wait wr_wait;
138 wq->sq.qid = c4iw_get_qpid(rdev, uctx);
142 wq->rq.qid = c4iw_get_qpid(rdev, uctx);
147 wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
152 wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
158 /* RQT must be a power of 2. */
159 wq->rq.rqt_size = roundup_pow_of_two(wq->rq.size);
160 wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
161 if (!wq->rq.rqt_hwaddr)
164 if (alloc_host_sq(rdev, &wq->sq))
167 memset(wq->sq.queue, 0, wq->sq.memsize);
168 pci_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
170 wq->rq.queue = contigmalloc(wq->rq.memsize,
171 M_DEVBUF, M_NOWAIT, 0ul, ~0ul, 4096, 0);
173 wq->rq.dma_addr = vtophys(wq->rq.queue);
177 "%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx", __func__,
178 wq->sq.queue, (unsigned long long)vtophys(wq->sq.queue),
179 wq->rq.queue, (unsigned long long)vtophys(wq->rq.queue));
180 memset(wq->rq.queue, 0, wq->rq.memsize);
181 pci_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
183 wq->db = (void *)((unsigned long)rman_get_virtual(sc->regs_res) +
184 MYPF_REG(SGE_PF_KDOORBELL));
185 wq->gts = (void *)((unsigned long)rman_get_virtual(rdev->adap->regs_res)
186 + MYPF_REG(SGE_PF_GTS));
188 wq->sq.udb = (u64)((char*)rman_get_virtual(rdev->adap->udbs_res) +
189 (wq->sq.qid << rdev->qpshift));
190 wq->sq.udb &= PAGE_MASK;
191 wq->rq.udb = (u64)((char*)rman_get_virtual(rdev->adap->udbs_res) +
192 (wq->rq.qid << rdev->qpshift));
193 wq->rq.udb &= PAGE_MASK;
198 /* build fw_ri_res_wr */
199 wr_len = sizeof *res_wr + 2 * sizeof *res;
201 wr = alloc_wrqe(wr_len, &sc->sge.mgmtq);
206 memset(res_wr, 0, wr_len);
207 res_wr->op_nres = cpu_to_be32(
208 V_FW_WR_OP(FW_RI_RES_WR) |
209 V_FW_RI_RES_WR_NRES(2) |
211 res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
212 res_wr->cookie = (unsigned long) &wr_wait;
214 res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
215 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
217 /* eqsize is the number of 64B entries plus the status page size. */
218 eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
219 (sc->params.sge.spg_len / EQ_ESIZE);
221 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
222 V_FW_RI_RES_WR_HOSTFCMODE(0) | /* no host cidx updates */
223 V_FW_RI_RES_WR_CPRIO(0) | /* don't keep in chip cache */
224 V_FW_RI_RES_WR_PCIECHN(0) | /* set by uP at ri_init time */
225 V_FW_RI_RES_WR_IQID(scq->cqid));
226 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
227 V_FW_RI_RES_WR_DCAEN(0) |
228 V_FW_RI_RES_WR_DCACPU(0) |
229 V_FW_RI_RES_WR_FBMIN(2) |
230 V_FW_RI_RES_WR_FBMAX(2) |
231 V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
232 V_FW_RI_RES_WR_CIDXFTHRESH(0) |
233 V_FW_RI_RES_WR_EQSIZE(eqsize));
234 res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
235 res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
237 res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
238 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
240 /* eqsize is the number of 64B entries plus the status page size. */
241 eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
242 (sc->params.sge.spg_len / EQ_ESIZE);
243 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
244 V_FW_RI_RES_WR_HOSTFCMODE(0) | /* no host cidx updates */
245 V_FW_RI_RES_WR_CPRIO(0) | /* don't keep in chip cache */
246 V_FW_RI_RES_WR_PCIECHN(0) | /* set by uP at ri_init time */
247 V_FW_RI_RES_WR_IQID(rcq->cqid));
248 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
249 V_FW_RI_RES_WR_DCAEN(0) |
250 V_FW_RI_RES_WR_DCACPU(0) |
251 V_FW_RI_RES_WR_FBMIN(2) |
252 V_FW_RI_RES_WR_FBMAX(2) |
253 V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
254 V_FW_RI_RES_WR_CIDXFTHRESH(0) |
255 V_FW_RI_RES_WR_EQSIZE(eqsize));
256 res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
257 res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
259 c4iw_init_wr_wait(&wr_wait);
262 ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
267 "%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx",
268 __func__, wq->sq.qid, wq->rq.qid, wq->db,
269 (unsigned long long)wq->sq.udb, (unsigned long long)wq->rq.udb);
273 contigfree(wq->rq.queue, wq->rq.memsize, M_DEVBUF);
275 dealloc_sq(rdev, &wq->sq);
277 c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
283 c4iw_put_qpid(rdev, wq->rq.qid, uctx);
285 c4iw_put_qpid(rdev, wq->sq.qid, uctx);
289 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
290 struct ib_send_wr *wr, int max, u32 *plenp)
297 dstp = (u8 *)immdp->data;
298 for (i = 0; i < wr->num_sge; i++) {
299 if ((plen + wr->sg_list[i].length) > max)
301 srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
302 plen += wr->sg_list[i].length;
303 rem = wr->sg_list[i].length;
305 if (dstp == (u8 *)&sq->queue[sq->size])
306 dstp = (u8 *)sq->queue;
307 if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
310 len = (u8 *)&sq->queue[sq->size] - dstp;
311 memcpy(dstp, srcp, len);
317 len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
319 memset(dstp, 0, len);
320 immdp->op = FW_RI_DATA_IMMD;
323 immdp->immdlen = cpu_to_be32(plen);
328 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
329 struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
330 int num_sge, u32 *plenp)
335 __be64 *flitp = (__be64 *)isglp->sge;
337 for (i = 0; i < num_sge; i++) {
338 if ((plen + sg_list[i].length) < plen)
340 plen += sg_list[i].length;
341 *flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
343 if (++flitp == queue_end)
345 *flitp = cpu_to_be64(sg_list[i].addr);
346 if (++flitp == queue_end)
349 *flitp = (__force __be64)0;
350 isglp->op = FW_RI_DATA_ISGL;
352 isglp->nsge = cpu_to_be16(num_sge);
359 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
360 struct ib_send_wr *wr, u8 *len16)
366 if (wr->num_sge > T4_MAX_SEND_SGE)
368 switch (wr->opcode) {
370 if (wr->send_flags & IB_SEND_SOLICITED)
371 wqe->send.sendop_pkd = cpu_to_be32(
372 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
374 wqe->send.sendop_pkd = cpu_to_be32(
375 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
376 wqe->send.stag_inv = 0;
378 case IB_WR_SEND_WITH_INV:
379 if (wr->send_flags & IB_SEND_SOLICITED)
380 wqe->send.sendop_pkd = cpu_to_be32(
381 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
383 wqe->send.sendop_pkd = cpu_to_be32(
384 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
385 wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
394 if (wr->send_flags & IB_SEND_INLINE) {
395 ret = build_immd(sq, wqe->send.u.immd_src, wr,
396 T4_MAX_SEND_INLINE, &plen);
399 size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
402 ret = build_isgl((__be64 *)sq->queue,
403 (__be64 *)&sq->queue[sq->size],
404 wqe->send.u.isgl_src,
405 wr->sg_list, wr->num_sge, &plen);
408 size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
409 wr->num_sge * sizeof(struct fw_ri_sge);
412 wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
413 wqe->send.u.immd_src[0].r1 = 0;
414 wqe->send.u.immd_src[0].r2 = 0;
415 wqe->send.u.immd_src[0].immdlen = 0;
416 size = sizeof wqe->send + sizeof(struct fw_ri_immd);
419 *len16 = DIV_ROUND_UP(size, 16);
420 wqe->send.plen = cpu_to_be32(plen);
424 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
425 struct ib_send_wr *wr, u8 *len16)
431 if (wr->num_sge > T4_MAX_SEND_SGE)
434 wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
435 wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
437 if (wr->send_flags & IB_SEND_INLINE) {
438 ret = build_immd(sq, wqe->write.u.immd_src, wr,
439 T4_MAX_WRITE_INLINE, &plen);
442 size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
445 ret = build_isgl((__be64 *)sq->queue,
446 (__be64 *)&sq->queue[sq->size],
447 wqe->write.u.isgl_src,
448 wr->sg_list, wr->num_sge, &plen);
451 size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
452 wr->num_sge * sizeof(struct fw_ri_sge);
455 wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
456 wqe->write.u.immd_src[0].r1 = 0;
457 wqe->write.u.immd_src[0].r2 = 0;
458 wqe->write.u.immd_src[0].immdlen = 0;
459 size = sizeof wqe->write + sizeof(struct fw_ri_immd);
462 *len16 = DIV_ROUND_UP(size, 16);
463 wqe->write.plen = cpu_to_be32(plen);
467 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
472 wqe->read.stag_src = cpu_to_be32(wr->wr.rdma.rkey);
473 wqe->read.to_src_hi = cpu_to_be32((u32)(wr->wr.rdma.remote_addr
475 wqe->read.to_src_lo = cpu_to_be32((u32)wr->wr.rdma.remote_addr);
476 wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
477 wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
478 wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
480 wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
482 wqe->read.stag_src = cpu_to_be32(2);
483 wqe->read.to_src_hi = 0;
484 wqe->read.to_src_lo = 0;
485 wqe->read.stag_sink = cpu_to_be32(2);
487 wqe->read.to_sink_hi = 0;
488 wqe->read.to_sink_lo = 0;
492 *len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
496 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
497 struct ib_recv_wr *wr, u8 *len16)
501 ret = build_isgl((__be64 *)qhp->wq.rq.queue,
502 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
503 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
506 *len16 = DIV_ROUND_UP(sizeof wqe->recv +
507 wr->num_sge * sizeof(struct fw_ri_sge), 16);
511 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
512 struct ib_send_wr *wr, u8 *len16)
515 struct fw_ri_immd *imdp;
518 int pbllen = roundup(wr->wr.fast_reg.page_list_len * sizeof(u64), 32);
521 if (wr->wr.fast_reg.page_list_len > T4_MAX_FR_DEPTH)
524 wqe->fr.qpbinde_to_dcacpu = 0;
525 wqe->fr.pgsz_shift = wr->wr.fast_reg.page_shift - 12;
526 wqe->fr.addr_type = FW_RI_VA_BASED_TO;
527 wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->wr.fast_reg.access_flags);
529 wqe->fr.len_lo = cpu_to_be32(wr->wr.fast_reg.length);
530 wqe->fr.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
531 wqe->fr.va_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
532 wqe->fr.va_lo_fbo = cpu_to_be32(wr->wr.fast_reg.iova_start &
534 WARN_ON(pbllen > T4_MAX_FR_IMMD);
535 imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
536 imdp->op = FW_RI_DATA_IMMD;
539 imdp->immdlen = cpu_to_be32(pbllen);
540 p = (__be64 *)(imdp + 1);
542 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
543 *p = cpu_to_be64((u64)wr->wr.fast_reg.page_list->page_list[i]);
545 if (++p == (__be64 *)&sq->queue[sq->size])
546 p = (__be64 *)sq->queue;
552 if (++p == (__be64 *)&sq->queue[sq->size])
553 p = (__be64 *)sq->queue;
555 *len16 = DIV_ROUND_UP(sizeof wqe->fr + sizeof *imdp + pbllen, 16);
559 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
562 wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
564 *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
568 void c4iw_qp_add_ref(struct ib_qp *qp)
570 CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
571 atomic_inc(&(to_c4iw_qp(qp)->refcnt));
574 void c4iw_qp_rem_ref(struct ib_qp *qp)
576 CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
577 if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
578 wake_up(&(to_c4iw_qp(qp)->wait));
581 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
582 struct ib_send_wr **bad_wr)
586 enum fw_wr_opcodes fw_opcode = 0;
587 enum fw_ri_wr_flags fw_flags;
591 struct t4_swsqe *swsqe;
595 qhp = to_c4iw_qp(ibqp);
596 spin_lock_irqsave(&qhp->lock, flag);
597 if (t4_wq_in_error(&qhp->wq)) {
598 spin_unlock_irqrestore(&qhp->lock, flag);
601 num_wrs = t4_sq_avail(&qhp->wq);
603 spin_unlock_irqrestore(&qhp->lock, flag);
612 wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
613 qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
616 if (wr->send_flags & IB_SEND_SOLICITED)
617 fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
618 if (wr->send_flags & IB_SEND_SIGNALED)
619 fw_flags |= FW_RI_COMPLETION_FLAG;
620 swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
621 switch (wr->opcode) {
622 case IB_WR_SEND_WITH_INV:
624 if (wr->send_flags & IB_SEND_FENCE)
625 fw_flags |= FW_RI_READ_FENCE_FLAG;
626 fw_opcode = FW_RI_SEND_WR;
627 if (wr->opcode == IB_WR_SEND)
628 swsqe->opcode = FW_RI_SEND;
630 swsqe->opcode = FW_RI_SEND_WITH_INV;
631 err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
633 case IB_WR_RDMA_WRITE:
634 fw_opcode = FW_RI_RDMA_WRITE_WR;
635 swsqe->opcode = FW_RI_RDMA_WRITE;
636 err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
638 case IB_WR_RDMA_READ:
639 case IB_WR_RDMA_READ_WITH_INV:
640 fw_opcode = FW_RI_RDMA_READ_WR;
641 swsqe->opcode = FW_RI_READ_REQ;
642 if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
643 fw_flags = FW_RI_RDMA_READ_INVALIDATE;
646 err = build_rdma_read(wqe, wr, &len16);
649 swsqe->read_len = wr->sg_list[0].length;
650 if (!qhp->wq.sq.oldest_read)
651 qhp->wq.sq.oldest_read = swsqe;
653 case IB_WR_FAST_REG_MR:
654 fw_opcode = FW_RI_FR_NSMR_WR;
655 swsqe->opcode = FW_RI_FAST_REGISTER;
656 err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16);
658 case IB_WR_LOCAL_INV:
659 if (wr->send_flags & IB_SEND_FENCE)
660 fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
661 fw_opcode = FW_RI_INV_LSTAG_WR;
662 swsqe->opcode = FW_RI_LOCAL_INV;
663 err = build_inv_stag(wqe, wr, &len16);
666 CTR2(KTR_IW_CXGBE, "%s post of type =%d TBD!", __func__,
674 swsqe->idx = qhp->wq.sq.pidx;
676 swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED);
677 swsqe->wr_id = wr->wr_id;
679 init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
682 "%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u",
683 __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
684 swsqe->opcode, swsqe->read_len);
687 t4_sq_produce(&qhp->wq, len16);
688 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
690 if (t4_wq_db_enabled(&qhp->wq))
691 t4_ring_sq_db(&qhp->wq, idx);
692 spin_unlock_irqrestore(&qhp->lock, flag);
696 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
697 struct ib_recv_wr **bad_wr)
701 union t4_recv_wr *wqe;
707 qhp = to_c4iw_qp(ibqp);
708 spin_lock_irqsave(&qhp->lock, flag);
709 if (t4_wq_in_error(&qhp->wq)) {
710 spin_unlock_irqrestore(&qhp->lock, flag);
713 num_wrs = t4_rq_avail(&qhp->wq);
715 spin_unlock_irqrestore(&qhp->lock, flag);
719 if (wr->num_sge > T4_MAX_RECV_SGE) {
724 wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
728 err = build_rdma_recv(qhp, wqe, wr, &len16);
736 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
738 wqe->recv.opcode = FW_RI_RECV_WR;
740 wqe->recv.wrid = qhp->wq.rq.pidx;
744 wqe->recv.len16 = len16;
745 CTR3(KTR_IW_CXGBE, "%s cookie 0x%llx pidx %u", __func__,
746 (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
747 t4_rq_produce(&qhp->wq, len16);
748 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
752 if (t4_wq_db_enabled(&qhp->wq))
753 t4_ring_rq_db(&qhp->wq, idx);
754 spin_unlock_irqrestore(&qhp->lock, flag);
758 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
763 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
773 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
778 status = CQE_STATUS(err_cqe);
779 opcode = CQE_OPCODE(err_cqe);
780 rqtype = RQ_TYPE(err_cqe);
781 send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
782 (opcode == FW_RI_SEND_WITH_SE_INV);
783 tagged = (opcode == FW_RI_RDMA_WRITE) ||
784 (rqtype && (opcode == FW_RI_READ_RESP));
789 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
790 *ecode = RDMAP_CANT_INV_STAG;
792 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
793 *ecode = RDMAP_INV_STAG;
797 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
798 if ((opcode == FW_RI_SEND_WITH_INV) ||
799 (opcode == FW_RI_SEND_WITH_SE_INV))
800 *ecode = RDMAP_CANT_INV_STAG;
802 *ecode = RDMAP_STAG_NOT_ASSOC;
805 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
806 *ecode = RDMAP_STAG_NOT_ASSOC;
809 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
810 *ecode = RDMAP_ACC_VIOL;
813 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
814 *ecode = RDMAP_TO_WRAP;
818 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
819 *ecode = DDPT_BASE_BOUNDS;
821 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
822 *ecode = RDMAP_BASE_BOUNDS;
825 case T4_ERR_INVALIDATE_SHARED_MR:
826 case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
827 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
828 *ecode = RDMAP_CANT_INV_STAG;
831 case T4_ERR_ECC_PSTAG:
832 case T4_ERR_INTERNAL_ERR:
833 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
836 case T4_ERR_OUT_OF_RQE:
837 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
838 *ecode = DDPU_INV_MSN_NOBUF;
840 case T4_ERR_PBL_ADDR_BOUND:
841 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
842 *ecode = DDPT_BASE_BOUNDS;
845 *layer_type = LAYER_MPA|DDP_LLP;
846 *ecode = MPA_CRC_ERR;
849 *layer_type = LAYER_MPA|DDP_LLP;
850 *ecode = MPA_MARKER_ERR;
852 case T4_ERR_PDU_LEN_ERR:
853 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
854 *ecode = DDPU_MSG_TOOBIG;
856 case T4_ERR_DDP_VERSION:
858 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
859 *ecode = DDPT_INV_VERS;
861 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
862 *ecode = DDPU_INV_VERS;
865 case T4_ERR_RDMA_VERSION:
866 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
867 *ecode = RDMAP_INV_VERS;
870 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
871 *ecode = RDMAP_INV_OPCODE;
873 case T4_ERR_DDP_QUEUE_NUM:
874 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
875 *ecode = DDPU_INV_QN;
879 case T4_ERR_MSN_RANGE:
880 case T4_ERR_IRD_OVERFLOW:
881 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
882 *ecode = DDPU_INV_MSN_RANGE;
885 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
889 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
890 *ecode = DDPU_INV_MO;
893 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
899 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
902 struct fw_ri_wr *wqe;
903 struct terminate_message *term;
905 struct socket *so = qhp->ep->com.so;
906 struct inpcb *inp = sotoinpcb(so);
907 struct tcpcb *tp = intotcpcb(inp);
908 struct toepcb *toep = tp->t_toe;
910 CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
911 qhp->wq.sq.qid, qhp->ep->hwtid);
913 wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
918 memset(wqe, 0, sizeof *wqe);
919 wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR));
920 wqe->flowid_len16 = cpu_to_be32(
921 V_FW_WR_FLOWID(qhp->ep->hwtid) |
922 V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
924 wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
925 wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
926 term = (struct terminate_message *)wqe->u.terminate.termmsg;
927 if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
928 term->layer_etype = qhp->attr.layer_etype;
929 term->ecode = qhp->attr.ecode;
931 build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
932 creds(toep, sizeof(*wqe));
933 t4_wrq_tx(qhp->rhp->rdev.adap, wr);
936 /* Assumes qhp lock is held. */
937 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
938 struct c4iw_cq *schp)
944 CTR4(KTR_IW_CXGBE, "%s qhp %p rchp %p schp %p", __func__, qhp, rchp,
947 /* locking hierarchy: cq lock first, then qp lock. */
948 spin_lock_irqsave(&rchp->lock, flag);
949 spin_lock(&qhp->lock);
950 c4iw_flush_hw_cq(&rchp->cq);
951 c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
952 flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
953 spin_unlock(&qhp->lock);
954 spin_unlock_irqrestore(&rchp->lock, flag);
956 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
957 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
958 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
961 /* locking hierarchy: cq lock first, then qp lock. */
962 spin_lock_irqsave(&schp->lock, flag);
963 spin_lock(&qhp->lock);
964 c4iw_flush_hw_cq(&schp->cq);
965 c4iw_count_scqes(&schp->cq, &qhp->wq, &count);
966 flushed = c4iw_flush_sq(&qhp->wq, &schp->cq, count);
967 spin_unlock(&qhp->lock);
968 spin_unlock_irqrestore(&schp->lock, flag);
970 spin_lock_irqsave(&schp->comp_handler_lock, flag);
971 (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
972 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
976 static void flush_qp(struct c4iw_qp *qhp)
978 struct c4iw_cq *rchp, *schp;
981 rchp = get_chp(qhp->rhp, qhp->attr.rcq);
982 schp = get_chp(qhp->rhp, qhp->attr.scq);
984 if (qhp->ibqp.uobject) {
985 t4_set_wq_in_error(&qhp->wq);
986 t4_set_cq_in_error(&rchp->cq);
987 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
988 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
989 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
991 t4_set_cq_in_error(&schp->cq);
992 spin_lock_irqsave(&schp->comp_handler_lock, flag);
993 (*schp->ibcq.comp_handler)(&schp->ibcq,
994 schp->ibcq.cq_context);
995 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
999 __flush_qp(qhp, rchp, schp);
1003 rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp, struct c4iw_ep *ep)
1005 struct c4iw_rdev *rdev = &rhp->rdev;
1006 struct adapter *sc = rdev->adap;
1007 struct fw_ri_wr *wqe;
1010 struct socket *so = ep->com.so;
1011 struct inpcb *inp = sotoinpcb(so);
1012 struct tcpcb *tp = intotcpcb(inp);
1013 struct toepcb *toep = tp->t_toe;
1015 KASSERT(rhp == qhp->rhp && ep == qhp->ep, ("%s: EDOOFUS", __func__));
1017 CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
1018 qhp->wq.sq.qid, ep->hwtid);
1020 wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
1025 memset(wqe, 0, sizeof *wqe);
1027 wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR) | F_FW_WR_COMPL);
1028 wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1029 V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1030 wqe->cookie = (unsigned long) &ep->com.wr_wait;
1031 wqe->u.fini.type = FW_RI_TYPE_FINI;
1033 c4iw_init_wr_wait(&ep->com.wr_wait);
1035 creds(toep, sizeof(*wqe));
1038 ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1039 qhp->wq.sq.qid, __func__);
1043 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1045 CTR2(KTR_IW_CXGBE, "%s p2p_type = %d", __func__, p2p_type);
1046 memset(&init->u, 0, sizeof init->u);
1048 case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1049 init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1050 init->u.write.stag_sink = cpu_to_be32(1);
1051 init->u.write.to_sink = cpu_to_be64(1);
1052 init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1053 init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1054 sizeof(struct fw_ri_immd),
1057 case FW_RI_INIT_P2PTYPE_READ_REQ:
1058 init->u.write.opcode = FW_RI_RDMA_READ_WR;
1059 init->u.read.stag_src = cpu_to_be32(1);
1060 init->u.read.to_src_lo = cpu_to_be32(1);
1061 init->u.read.stag_sink = cpu_to_be32(1);
1062 init->u.read.to_sink_lo = cpu_to_be32(1);
1063 init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1069 creds(struct toepcb *toep, size_t wrsize)
1071 struct ofld_tx_sdesc *txsd;
1073 CTR3(KTR_IW_CXGBE, "%s:creB %p %u", __func__, toep , wrsize);
1074 INP_WLOCK(toep->inp);
1075 txsd = &toep->txsd[toep->txsd_pidx];
1076 txsd->tx_credits = howmany(wrsize, 16);
1078 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
1079 ("%s: not enough credits (%d)", __func__, toep->tx_credits));
1080 toep->tx_credits -= txsd->tx_credits;
1081 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1082 toep->txsd_pidx = 0;
1084 INP_WUNLOCK(toep->inp);
1085 CTR5(KTR_IW_CXGBE, "%s:creE %p %u %u %u", __func__, toep ,
1086 txsd->tx_credits, toep->tx_credits, toep->txsd_pidx);
1089 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1091 struct fw_ri_wr *wqe;
1094 struct c4iw_ep *ep = qhp->ep;
1095 struct c4iw_rdev *rdev = &qhp->rhp->rdev;
1096 struct adapter *sc = rdev->adap;
1097 struct socket *so = ep->com.so;
1098 struct inpcb *inp = sotoinpcb(so);
1099 struct tcpcb *tp = intotcpcb(inp);
1100 struct toepcb *toep = tp->t_toe;
1102 CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
1103 qhp->wq.sq.qid, ep->hwtid);
1105 wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
1110 memset(wqe, 0, sizeof *wqe);
1112 wqe->op_compl = cpu_to_be32(
1113 V_FW_WR_OP(FW_RI_WR) |
1115 wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1116 V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1118 wqe->cookie = (unsigned long) &ep->com.wr_wait;
1120 wqe->u.init.type = FW_RI_TYPE_INIT;
1121 wqe->u.init.mpareqbit_p2ptype =
1122 V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1123 V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1124 wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1125 if (qhp->attr.mpa_attr.recv_marker_enabled)
1126 wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1127 if (qhp->attr.mpa_attr.xmit_marker_enabled)
1128 wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1129 if (qhp->attr.mpa_attr.crc_enabled)
1130 wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1132 wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1133 FW_RI_QP_RDMA_WRITE_ENABLE |
1134 FW_RI_QP_BIND_ENABLE;
1135 if (!qhp->ibqp.uobject)
1136 wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1137 FW_RI_QP_STAG0_ENABLE;
1138 wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1139 wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1140 wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1141 wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1142 wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1143 wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1144 wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1145 wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1146 wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1147 wqe->u.init.iss = cpu_to_be32(ep->snd_seq);
1148 wqe->u.init.irs = cpu_to_be32(ep->rcv_seq);
1149 wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1150 wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1152 if (qhp->attr.mpa_attr.initiator)
1153 build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1155 c4iw_init_wr_wait(&ep->com.wr_wait);
1157 creds(toep, sizeof(*wqe));
1160 ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1161 qhp->wq.sq.qid, __func__);
1163 toep->ulp_mode = ULP_MODE_RDMA;
1168 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1169 enum c4iw_qp_attr_mask mask,
1170 struct c4iw_qp_attributes *attrs,
1174 struct c4iw_qp_attributes newattr = qhp->attr;
1179 struct c4iw_ep *ep = NULL;
1181 CTR5(KTR_IW_CXGBE, "%s qhp %p sqid 0x%x rqid 0x%x ep %p", __func__, qhp,
1182 qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep);
1183 CTR3(KTR_IW_CXGBE, "%s state %d -> %d", __func__, qhp->attr.state,
1184 (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1186 mutex_lock(&qhp->mutex);
1188 /* Process attr changes if in IDLE */
1189 if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1190 if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1194 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1195 newattr.enable_rdma_read = attrs->enable_rdma_read;
1196 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1197 newattr.enable_rdma_write = attrs->enable_rdma_write;
1198 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1199 newattr.enable_bind = attrs->enable_bind;
1200 if (mask & C4IW_QP_ATTR_MAX_ORD) {
1201 if (attrs->max_ord > c4iw_max_read_depth) {
1205 newattr.max_ord = attrs->max_ord;
1207 if (mask & C4IW_QP_ATTR_MAX_IRD) {
1208 if (attrs->max_ird > c4iw_max_read_depth) {
1212 newattr.max_ird = attrs->max_ird;
1214 qhp->attr = newattr;
1217 if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1219 if (qhp->attr.state == attrs->next_state)
1222 switch (qhp->attr.state) {
1223 case C4IW_QP_STATE_IDLE:
1224 switch (attrs->next_state) {
1225 case C4IW_QP_STATE_RTS:
1226 if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1230 if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1234 qhp->attr.mpa_attr = attrs->mpa_attr;
1235 qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1236 qhp->ep = qhp->attr.llp_stream_handle;
1237 set_state(qhp, C4IW_QP_STATE_RTS);
1240 * Ref the endpoint here and deref when we
1241 * disassociate the endpoint from the QP. This
1242 * happens in CLOSING->IDLE transition or *->ERROR
1245 c4iw_get_ep(&qhp->ep->com);
1246 ret = rdma_init(rhp, qhp);
1250 case C4IW_QP_STATE_ERROR:
1251 set_state(qhp, C4IW_QP_STATE_ERROR);
1259 case C4IW_QP_STATE_RTS:
1260 switch (attrs->next_state) {
1261 case C4IW_QP_STATE_CLOSING:
1262 //Fixme: Use atomic_read as same as Linux
1263 BUG_ON(qhp->ep->com.kref.count < 2);
1264 set_state(qhp, C4IW_QP_STATE_CLOSING);
1269 c4iw_get_ep(&qhp->ep->com);
1271 if (qhp->ibqp.uobject)
1272 t4_set_wq_in_error(&qhp->wq);
1273 ret = rdma_fini(rhp, qhp, ep);
1277 case C4IW_QP_STATE_TERMINATE:
1278 set_state(qhp, C4IW_QP_STATE_TERMINATE);
1279 qhp->attr.layer_etype = attrs->layer_etype;
1280 qhp->attr.ecode = attrs->ecode;
1281 if (qhp->ibqp.uobject)
1282 t4_set_wq_in_error(&qhp->wq);
1287 c4iw_get_ep(&qhp->ep->com);
1289 case C4IW_QP_STATE_ERROR:
1290 set_state(qhp, C4IW_QP_STATE_ERROR);
1291 if (qhp->ibqp.uobject)
1292 t4_set_wq_in_error(&qhp->wq);
1297 c4iw_get_ep(&qhp->ep->com);
1306 case C4IW_QP_STATE_CLOSING:
1311 switch (attrs->next_state) {
1312 case C4IW_QP_STATE_IDLE:
1314 set_state(qhp, C4IW_QP_STATE_IDLE);
1315 qhp->attr.llp_stream_handle = NULL;
1316 c4iw_put_ep(&qhp->ep->com);
1318 wake_up(&qhp->wait);
1320 case C4IW_QP_STATE_ERROR:
1327 case C4IW_QP_STATE_ERROR:
1328 if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1332 if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1336 set_state(qhp, C4IW_QP_STATE_IDLE);
1338 case C4IW_QP_STATE_TERMINATE:
1346 printf("%s in a bad state %d\n",
1347 __func__, qhp->attr.state);
1354 CTR3(KTR_IW_CXGBE, "%s disassociating ep %p qpid 0x%x", __func__,
1355 qhp->ep, qhp->wq.sq.qid);
1357 /* disassociate the LLP connection */
1358 qhp->attr.llp_stream_handle = NULL;
1362 set_state(qhp, C4IW_QP_STATE_ERROR);
1366 wake_up(&qhp->wait);
1368 mutex_unlock(&qhp->mutex);
1371 post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1374 * If disconnect is 1, then we need to initiate a disconnect
1375 * on the EP. This can be a normal close (RTS->CLOSING) or
1376 * an abnormal close (RTS/CLOSING->ERROR).
1379 c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1381 c4iw_put_ep(&ep->com);
1385 * If free is 1, then we've disassociated the EP from the QP
1386 * and we need to dereference the EP.
1389 c4iw_put_ep(&ep->com);
1390 CTR2(KTR_IW_CXGBE, "%s exit state %d", __func__, qhp->attr.state);
1394 static int enable_qp_db(int id, void *p, void *data)
1396 struct c4iw_qp *qp = p;
1398 t4_enable_wq_db(&qp->wq);
1402 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1404 struct c4iw_dev *rhp;
1405 struct c4iw_qp *qhp;
1406 struct c4iw_qp_attributes attrs;
1407 struct c4iw_ucontext *ucontext;
1409 CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ib_qp);
1410 qhp = to_c4iw_qp(ib_qp);
1413 attrs.next_state = C4IW_QP_STATE_ERROR;
1414 if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1415 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1417 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1418 wait_event(qhp->wait, !qhp->ep);
1420 spin_lock_irq(&rhp->lock);
1421 remove_handle_nolock(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1423 BUG_ON(rhp->qpcnt < 0);
1424 if (rhp->qpcnt <= db_fc_threshold && rhp->db_state == FLOW_CONTROL) {
1425 rhp->rdev.stats.db_state_transitions++;
1426 rhp->db_state = NORMAL;
1427 idr_for_each(&rhp->qpidr, enable_qp_db, NULL);
1429 spin_unlock_irq(&rhp->lock);
1430 atomic_dec(&qhp->refcnt);
1431 wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1433 ucontext = ib_qp->uobject ?
1434 to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1435 destroy_qp(&rhp->rdev, &qhp->wq,
1436 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1438 CTR3(KTR_IW_CXGBE, "%s ib_qp %p qpid 0x%0x", __func__, ib_qp,
1444 static int disable_qp_db(int id, void *p, void *data)
1446 struct c4iw_qp *qp = p;
1448 t4_disable_wq_db(&qp->wq);
1453 c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1454 struct ib_udata *udata)
1456 struct c4iw_dev *rhp;
1457 struct c4iw_qp *qhp;
1458 struct c4iw_pd *php;
1459 struct c4iw_cq *schp;
1460 struct c4iw_cq *rchp;
1461 struct c4iw_create_qp_resp uresp;
1463 struct c4iw_ucontext *ucontext;
1465 struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4;
1467 CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
1469 if (attrs->qp_type != IB_QPT_RC)
1470 return ERR_PTR(-EINVAL);
1472 php = to_c4iw_pd(pd);
1474 schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1475 rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1477 return ERR_PTR(-EINVAL);
1479 if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1480 return ERR_PTR(-EINVAL);
1482 rqsize = roundup(attrs->cap.max_recv_wr + 1, 16);
1483 if (rqsize > T4_MAX_RQ_SIZE)
1484 return ERR_PTR(-E2BIG);
1486 sqsize = roundup(attrs->cap.max_send_wr + 1, 16);
1487 if (sqsize > T4_MAX_SQ_SIZE)
1488 return ERR_PTR(-E2BIG);
1490 ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1493 qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1495 return ERR_PTR(-ENOMEM);
1496 qhp->wq.sq.size = sqsize;
1497 qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
1498 qhp->wq.rq.size = rqsize;
1499 qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
1502 qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1503 qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1506 CTR5(KTR_IW_CXGBE, "%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu",
1507 __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1509 ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1510 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1514 attrs->cap.max_recv_wr = rqsize - 1;
1515 attrs->cap.max_send_wr = sqsize - 1;
1516 attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1519 qhp->attr.pd = php->pdid;
1520 qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1521 qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1522 qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1523 qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1524 qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1525 qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1526 qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1527 qhp->attr.state = C4IW_QP_STATE_IDLE;
1528 qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1529 qhp->attr.enable_rdma_read = 1;
1530 qhp->attr.enable_rdma_write = 1;
1531 qhp->attr.enable_bind = 1;
1532 qhp->attr.max_ord = 1;
1533 qhp->attr.max_ird = 1;
1534 spin_lock_init(&qhp->lock);
1535 mutex_init(&qhp->mutex);
1536 init_waitqueue_head(&qhp->wait);
1537 atomic_set(&qhp->refcnt, 1);
1539 spin_lock_irq(&rhp->lock);
1540 if (rhp->db_state != NORMAL)
1541 t4_disable_wq_db(&qhp->wq);
1542 if (++rhp->qpcnt > db_fc_threshold && rhp->db_state == NORMAL) {
1543 rhp->rdev.stats.db_state_transitions++;
1544 rhp->db_state = FLOW_CONTROL;
1545 idr_for_each(&rhp->qpidr, disable_qp_db, NULL);
1547 ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1548 spin_unlock_irq(&rhp->lock);
1553 mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1558 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1563 mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1568 mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1574 uresp.qid_mask = rhp->rdev.qpmask;
1575 uresp.sqid = qhp->wq.sq.qid;
1576 uresp.sq_size = qhp->wq.sq.size;
1577 uresp.sq_memsize = qhp->wq.sq.memsize;
1578 uresp.rqid = qhp->wq.rq.qid;
1579 uresp.rq_size = qhp->wq.rq.size;
1580 uresp.rq_memsize = qhp->wq.rq.memsize;
1581 spin_lock(&ucontext->mmap_lock);
1582 uresp.sq_key = ucontext->key;
1583 ucontext->key += PAGE_SIZE;
1584 uresp.rq_key = ucontext->key;
1585 ucontext->key += PAGE_SIZE;
1586 uresp.sq_db_gts_key = ucontext->key;
1587 ucontext->key += PAGE_SIZE;
1588 uresp.rq_db_gts_key = ucontext->key;
1589 ucontext->key += PAGE_SIZE;
1590 spin_unlock(&ucontext->mmap_lock);
1591 ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1594 mm1->key = uresp.sq_key;
1595 mm1->addr = qhp->wq.sq.phys_addr;
1596 mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1597 CTR4(KTR_IW_CXGBE, "%s mm1 %x, %x, %d", __func__, mm1->key,
1598 mm1->addr, mm1->len);
1599 insert_mmap(ucontext, mm1);
1600 mm2->key = uresp.rq_key;
1601 mm2->addr = vtophys(qhp->wq.rq.queue);
1602 mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1603 CTR4(KTR_IW_CXGBE, "%s mm2 %x, %x, %d", __func__, mm2->key,
1604 mm2->addr, mm2->len);
1605 insert_mmap(ucontext, mm2);
1606 mm3->key = uresp.sq_db_gts_key;
1607 mm3->addr = qhp->wq.sq.udb;
1608 mm3->len = PAGE_SIZE;
1609 CTR4(KTR_IW_CXGBE, "%s mm3 %x, %x, %d", __func__, mm3->key,
1610 mm3->addr, mm3->len);
1611 insert_mmap(ucontext, mm3);
1612 mm4->key = uresp.rq_db_gts_key;
1613 mm4->addr = qhp->wq.rq.udb;
1614 mm4->len = PAGE_SIZE;
1615 CTR4(KTR_IW_CXGBE, "%s mm4 %x, %x, %d", __func__, mm4->key,
1616 mm4->addr, mm4->len);
1617 insert_mmap(ucontext, mm4);
1619 qhp->ibqp.qp_num = qhp->wq.sq.qid;
1620 init_timer(&(qhp->timer));
1622 "%s qhp %p sq_num_entries %d, rq_num_entries %d qpid 0x%0x",
1623 __func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
1635 remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1637 destroy_qp(&rhp->rdev, &qhp->wq,
1638 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1641 return ERR_PTR(ret);
1644 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1645 int attr_mask, struct ib_udata *udata)
1647 struct c4iw_dev *rhp;
1648 struct c4iw_qp *qhp;
1649 enum c4iw_qp_attr_mask mask = 0;
1650 struct c4iw_qp_attributes attrs;
1652 CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ibqp);
1654 /* iwarp does not support the RTR state */
1655 if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1656 attr_mask &= ~IB_QP_STATE;
1658 /* Make sure we still have something left to do */
1662 memset(&attrs, 0, sizeof attrs);
1663 qhp = to_c4iw_qp(ibqp);
1666 attrs.next_state = c4iw_convert_state(attr->qp_state);
1667 attrs.enable_rdma_read = (attr->qp_access_flags &
1668 IB_ACCESS_REMOTE_READ) ? 1 : 0;
1669 attrs.enable_rdma_write = (attr->qp_access_flags &
1670 IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1671 attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1674 mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1675 mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1676 (C4IW_QP_ATTR_ENABLE_RDMA_READ |
1677 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1678 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1681 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1682 * ringing the queue db when we're in DB_FULL mode.
1684 attrs.sq_db_inc = attr->sq_psn;
1685 attrs.rq_db_inc = attr->rq_psn;
1686 mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1687 mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1689 return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1692 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1694 CTR3(KTR_IW_CXGBE, "%s ib_dev %p qpn 0x%x", __func__, dev, qpn);
1695 return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1698 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1699 int attr_mask, struct ib_qp_init_attr *init_attr)
1701 struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1703 memset(attr, 0, sizeof *attr);
1704 memset(init_attr, 0, sizeof *init_attr);
1705 attr->qp_state = to_ib_qp_state(qhp->attr.state);