2 * Copyright (c) 2009-2010 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
33 #include <linux/module.h>
37 static int db_delay_usecs = 1;
38 module_param(db_delay_usecs, int, 0644);
39 MODULE_PARM_DESC(db_delay_usecs, "Usecs to delay awaiting db fifo to drain");
41 static int ocqp_support = 1;
42 module_param(ocqp_support, int, 0644);
43 MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=1)");
45 int db_fc_threshold = 1000;
46 module_param(db_fc_threshold, int, 0644);
47 MODULE_PARM_DESC(db_fc_threshold,
48 "QP count/threshold that triggers"
49 " automatic db flow control mode (default = 1000)");
51 int db_coalescing_threshold;
52 module_param(db_coalescing_threshold, int, 0644);
53 MODULE_PARM_DESC(db_coalescing_threshold,
54 "QP count/threshold that triggers"
55 " disabling db coalescing (default = 0)");
57 static int max_fr_immd = T4_MAX_FR_IMMD;
58 module_param(max_fr_immd, int, 0644);
59 MODULE_PARM_DESC(max_fr_immd, "fastreg threshold for using DSGL instead of immedate");
61 static int alloc_ird(struct c4iw_dev *dev, u32 ird)
65 spin_lock_irq(&dev->lock);
66 if (ird <= dev->avail_ird)
67 dev->avail_ird -= ird;
70 spin_unlock_irq(&dev->lock);
73 dev_warn(&dev->rdev.lldi.pdev->dev,
74 "device IRD resources exhausted\n");
79 static void free_ird(struct c4iw_dev *dev, int ird)
81 spin_lock_irq(&dev->lock);
82 dev->avail_ird += ird;
83 spin_unlock_irq(&dev->lock);
86 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
89 spin_lock_irqsave(&qhp->lock, flag);
90 qhp->attr.state = state;
91 spin_unlock_irqrestore(&qhp->lock, flag);
94 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
96 c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
99 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
101 dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
102 pci_unmap_addr(sq, mapping));
105 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
107 if (t4_sq_onchip(sq))
108 dealloc_oc_sq(rdev, sq);
110 dealloc_host_sq(rdev, sq);
113 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
115 if (!ocqp_support || !ocqp_supported(&rdev->lldi))
117 sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
120 sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
121 rdev->lldi.vr->ocq.start;
122 sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
123 rdev->lldi.vr->ocq.start);
124 sq->flags |= T4_SQ_ONCHIP;
128 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
130 sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
131 &(sq->dma_addr), GFP_KERNEL);
134 sq->phys_addr = virt_to_phys(sq->queue);
135 pci_unmap_addr_set(sq, mapping, sq->dma_addr);
139 static int alloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq, int user)
143 ret = alloc_oc_sq(rdev, sq);
145 ret = alloc_host_sq(rdev, sq);
149 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
150 struct c4iw_dev_ucontext *uctx)
153 * uP clears EQ contexts when the connection exits rdma mode,
154 * so no need to post a RESET WR for these EQs.
156 dma_free_coherent(&(rdev->lldi.pdev->dev),
157 wq->rq.memsize, wq->rq.queue,
158 dma_unmap_addr(&wq->rq, mapping));
159 dealloc_sq(rdev, &wq->sq);
160 c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
163 c4iw_put_qpid(rdev, wq->rq.qid, uctx);
164 c4iw_put_qpid(rdev, wq->sq.qid, uctx);
169 * Determine the BAR2 virtual address and qid. If pbar2_pa is not NULL,
170 * then this is a user mapping so compute the page-aligned physical address
173 void __iomem *c4iw_bar2_addrs(struct c4iw_rdev *rdev, unsigned int qid,
174 enum cxgb4_bar2_qtype qtype,
175 unsigned int *pbar2_qid, u64 *pbar2_pa)
180 ret = cxgb4_bar2_sge_qregs(rdev->lldi.ports[0], qid, qtype,
182 &bar2_qoffset, pbar2_qid);
187 *pbar2_pa = (rdev->bar2_pa + bar2_qoffset) & PAGE_MASK;
189 if (is_t4(rdev->lldi.adapter_type))
192 return rdev->bar2_kva + bar2_qoffset;
195 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
196 struct t4_cq *rcq, struct t4_cq *scq,
197 struct c4iw_dev_ucontext *uctx)
199 int user = (uctx != &rdev->uctx);
200 struct fw_ri_res_wr *res_wr;
201 struct fw_ri_res *res;
203 struct c4iw_wr_wait wr_wait;
208 wq->sq.qid = c4iw_get_qpid(rdev, uctx);
212 wq->rq.qid = c4iw_get_qpid(rdev, uctx);
219 wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
226 wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
235 * RQT must be a power of 2 and at least 16 deep.
237 wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
238 wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
239 if (!wq->rq.rqt_hwaddr) {
244 ret = alloc_sq(rdev, &wq->sq, user);
247 memset(wq->sq.queue, 0, wq->sq.memsize);
248 dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
250 wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
251 wq->rq.memsize, &(wq->rq.dma_addr),
257 PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
258 __func__, wq->sq.queue,
259 (unsigned long long)virt_to_phys(wq->sq.queue),
261 (unsigned long long)virt_to_phys(wq->rq.queue));
262 memset(wq->rq.queue, 0, wq->rq.memsize);
263 dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
265 wq->db = rdev->lldi.db_reg;
267 wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid, T4_BAR2_QTYPE_EGRESS,
269 user ? &wq->sq.bar2_pa : NULL);
270 wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid, T4_BAR2_QTYPE_EGRESS,
272 user ? &wq->rq.bar2_pa : NULL);
275 * User mode must have bar2 access.
277 if (user && (!wq->sq.bar2_pa || !wq->rq.bar2_pa)) {
278 pr_warn(MOD "%s: sqid %u or rqid %u not in BAR2 range.\n",
279 pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
287 /* build fw_ri_res_wr */
288 wr_len = sizeof *res_wr + 2 * sizeof *res;
290 skb = alloc_skb(wr_len, GFP_KERNEL);
295 set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
297 res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
298 memset(res_wr, 0, wr_len);
299 res_wr->op_nres = cpu_to_be32(
300 FW_WR_OP_V(FW_RI_RES_WR) |
301 FW_RI_RES_WR_NRES_V(2) |
303 res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
304 res_wr->cookie = (uintptr_t)&wr_wait;
306 res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
307 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
310 * eqsize is the number of 64B entries plus the status page size.
312 eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
313 rdev->hw_queue.t4_eq_status_entries;
315 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
316 FW_RI_RES_WR_HOSTFCMODE_V(0) | /* no host cidx updates */
317 FW_RI_RES_WR_CPRIO_V(0) | /* don't keep in chip cache */
318 FW_RI_RES_WR_PCIECHN_V(0) | /* set by uP at ri_init time */
319 (t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
320 FW_RI_RES_WR_IQID_V(scq->cqid));
321 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
322 FW_RI_RES_WR_DCAEN_V(0) |
323 FW_RI_RES_WR_DCACPU_V(0) |
324 FW_RI_RES_WR_FBMIN_V(2) |
325 FW_RI_RES_WR_FBMAX_V(2) |
326 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
327 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
328 FW_RI_RES_WR_EQSIZE_V(eqsize));
329 res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
330 res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
332 res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
333 res->u.sqrq.op = FW_RI_RES_OP_WRITE;
336 * eqsize is the number of 64B entries plus the status page size.
338 eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
339 rdev->hw_queue.t4_eq_status_entries;
340 res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
341 FW_RI_RES_WR_HOSTFCMODE_V(0) | /* no host cidx updates */
342 FW_RI_RES_WR_CPRIO_V(0) | /* don't keep in chip cache */
343 FW_RI_RES_WR_PCIECHN_V(0) | /* set by uP at ri_init time */
344 FW_RI_RES_WR_IQID_V(rcq->cqid));
345 res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
346 FW_RI_RES_WR_DCAEN_V(0) |
347 FW_RI_RES_WR_DCACPU_V(0) |
348 FW_RI_RES_WR_FBMIN_V(2) |
349 FW_RI_RES_WR_FBMAX_V(2) |
350 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
351 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
352 FW_RI_RES_WR_EQSIZE_V(eqsize));
353 res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
354 res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
356 c4iw_init_wr_wait(&wr_wait);
358 ret = c4iw_ofld_send(rdev, skb);
361 ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
365 PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
366 __func__, wq->sq.qid, wq->rq.qid, wq->db,
367 wq->sq.bar2_va, wq->rq.bar2_va);
371 dma_free_coherent(&(rdev->lldi.pdev->dev),
372 wq->rq.memsize, wq->rq.queue,
373 dma_unmap_addr(&wq->rq, mapping));
375 dealloc_sq(rdev, &wq->sq);
377 c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
383 c4iw_put_qpid(rdev, wq->rq.qid, uctx);
385 c4iw_put_qpid(rdev, wq->sq.qid, uctx);
389 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
390 struct ib_send_wr *wr, int max, u32 *plenp)
397 dstp = (u8 *)immdp->data;
398 for (i = 0; i < wr->num_sge; i++) {
399 if ((plen + wr->sg_list[i].length) > max)
401 srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
402 plen += wr->sg_list[i].length;
403 rem = wr->sg_list[i].length;
405 if (dstp == (u8 *)&sq->queue[sq->size])
406 dstp = (u8 *)sq->queue;
407 if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
410 len = (u8 *)&sq->queue[sq->size] - dstp;
411 memcpy(dstp, srcp, len);
417 len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
419 memset(dstp, 0, len);
420 immdp->op = FW_RI_DATA_IMMD;
423 immdp->immdlen = cpu_to_be32(plen);
428 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
429 struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
430 int num_sge, u32 *plenp)
435 __be64 *flitp = (__be64 *)isglp->sge;
437 for (i = 0; i < num_sge; i++) {
438 if ((plen + sg_list[i].length) < plen)
440 plen += sg_list[i].length;
441 *flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
443 if (++flitp == queue_end)
445 *flitp = cpu_to_be64(sg_list[i].addr);
446 if (++flitp == queue_end)
449 *flitp = (__force __be64)0;
450 isglp->op = FW_RI_DATA_ISGL;
452 isglp->nsge = cpu_to_be16(num_sge);
459 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
460 struct ib_send_wr *wr, u8 *len16)
466 if (wr->num_sge > T4_MAX_SEND_SGE)
468 switch (wr->opcode) {
470 if (wr->send_flags & IB_SEND_SOLICITED)
471 wqe->send.sendop_pkd = cpu_to_be32(
472 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
474 wqe->send.sendop_pkd = cpu_to_be32(
475 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
476 wqe->send.stag_inv = 0;
478 case IB_WR_SEND_WITH_INV:
479 if (wr->send_flags & IB_SEND_SOLICITED)
480 wqe->send.sendop_pkd = cpu_to_be32(
481 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE_INV));
483 wqe->send.sendop_pkd = cpu_to_be32(
484 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_INV));
485 wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
496 if (wr->send_flags & IB_SEND_INLINE) {
497 ret = build_immd(sq, wqe->send.u.immd_src, wr,
498 T4_MAX_SEND_INLINE, &plen);
501 size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
504 ret = build_isgl((__be64 *)sq->queue,
505 (__be64 *)&sq->queue[sq->size],
506 wqe->send.u.isgl_src,
507 wr->sg_list, wr->num_sge, &plen);
510 size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
511 wr->num_sge * sizeof(struct fw_ri_sge);
514 wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
515 wqe->send.u.immd_src[0].r1 = 0;
516 wqe->send.u.immd_src[0].r2 = 0;
517 wqe->send.u.immd_src[0].immdlen = 0;
518 size = sizeof wqe->send + sizeof(struct fw_ri_immd);
521 *len16 = DIV_ROUND_UP(size, 16);
522 wqe->send.plen = cpu_to_be32(plen);
526 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
527 struct ib_send_wr *wr, u8 *len16)
533 if (wr->num_sge > T4_MAX_SEND_SGE)
536 wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
537 wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
539 if (wr->send_flags & IB_SEND_INLINE) {
540 ret = build_immd(sq, wqe->write.u.immd_src, wr,
541 T4_MAX_WRITE_INLINE, &plen);
544 size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
547 ret = build_isgl((__be64 *)sq->queue,
548 (__be64 *)&sq->queue[sq->size],
549 wqe->write.u.isgl_src,
550 wr->sg_list, wr->num_sge, &plen);
553 size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
554 wr->num_sge * sizeof(struct fw_ri_sge);
557 wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
558 wqe->write.u.immd_src[0].r1 = 0;
559 wqe->write.u.immd_src[0].r2 = 0;
560 wqe->write.u.immd_src[0].immdlen = 0;
561 size = sizeof wqe->write + sizeof(struct fw_ri_immd);
564 *len16 = DIV_ROUND_UP(size, 16);
565 wqe->write.plen = cpu_to_be32(plen);
569 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
574 wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
575 wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
577 wqe->read.to_src_lo = cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
578 wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
579 wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
580 wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
582 wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
584 wqe->read.stag_src = cpu_to_be32(2);
585 wqe->read.to_src_hi = 0;
586 wqe->read.to_src_lo = 0;
587 wqe->read.stag_sink = cpu_to_be32(2);
589 wqe->read.to_sink_hi = 0;
590 wqe->read.to_sink_lo = 0;
594 *len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
598 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
599 struct ib_recv_wr *wr, u8 *len16)
603 ret = build_isgl((__be64 *)qhp->wq.rq.queue,
604 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
605 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
608 *len16 = DIV_ROUND_UP(sizeof wqe->recv +
609 wr->num_sge * sizeof(struct fw_ri_sge), 16);
613 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
614 struct ib_reg_wr *wr, u8 *len16, u8 t5dev)
616 struct c4iw_mr *mhp = to_c4iw_mr(wr->mr);
617 struct fw_ri_immd *imdp;
620 int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
623 if (mhp->mpl_len > t4_max_fr_depth(use_dsgl))
626 wqe->fr.qpbinde_to_dcacpu = 0;
627 wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
628 wqe->fr.addr_type = FW_RI_VA_BASED_TO;
629 wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
631 wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
632 wqe->fr.stag = cpu_to_be32(wr->key);
633 wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
634 wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
637 if (t5dev && use_dsgl && (pbllen > max_fr_immd)) {
638 struct fw_ri_dsgl *sglp;
640 for (i = 0; i < mhp->mpl_len; i++)
641 mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
643 sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
644 sglp->op = FW_RI_DATA_DSGL;
646 sglp->nsge = cpu_to_be16(1);
647 sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
648 sglp->len0 = cpu_to_be32(pbllen);
650 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
652 imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
653 imdp->op = FW_RI_DATA_IMMD;
656 imdp->immdlen = cpu_to_be32(pbllen);
657 p = (__be64 *)(imdp + 1);
659 for (i = 0; i < mhp->mpl_len; i++) {
660 *p = cpu_to_be64((u64)mhp->mpl[i]);
662 if (++p == (__be64 *)&sq->queue[sq->size])
663 p = (__be64 *)sq->queue;
669 if (++p == (__be64 *)&sq->queue[sq->size])
670 p = (__be64 *)sq->queue;
672 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
678 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
681 wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
683 *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
687 void c4iw_qp_add_ref(struct ib_qp *qp)
689 PDBG("%s ib_qp %p\n", __func__, qp);
690 atomic_inc(&(to_c4iw_qp(qp)->refcnt));
693 void c4iw_qp_rem_ref(struct ib_qp *qp)
695 PDBG("%s ib_qp %p\n", __func__, qp);
696 if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
697 wake_up(&(to_c4iw_qp(qp)->wait));
700 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
702 if (list_empty(entry))
703 list_add_tail(entry, head);
706 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
710 spin_lock_irqsave(&qhp->rhp->lock, flags);
711 spin_lock(&qhp->lock);
712 if (qhp->rhp->db_state == NORMAL)
713 t4_ring_sq_db(&qhp->wq, inc, NULL);
715 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
716 qhp->wq.sq.wq_pidx_inc += inc;
718 spin_unlock(&qhp->lock);
719 spin_unlock_irqrestore(&qhp->rhp->lock, flags);
723 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
727 spin_lock_irqsave(&qhp->rhp->lock, flags);
728 spin_lock(&qhp->lock);
729 if (qhp->rhp->db_state == NORMAL)
730 t4_ring_rq_db(&qhp->wq, inc, NULL);
732 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
733 qhp->wq.rq.wq_pidx_inc += inc;
735 spin_unlock(&qhp->lock);
736 spin_unlock_irqrestore(&qhp->rhp->lock, flags);
740 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
741 struct ib_send_wr **bad_wr)
745 enum fw_wr_opcodes fw_opcode = 0;
746 enum fw_ri_wr_flags fw_flags;
748 union t4_wr *wqe = NULL;
750 struct t4_swsqe *swsqe;
754 qhp = to_c4iw_qp(ibqp);
755 spin_lock_irqsave(&qhp->lock, flag);
756 if (t4_wq_in_error(&qhp->wq)) {
757 spin_unlock_irqrestore(&qhp->lock, flag);
760 num_wrs = t4_sq_avail(&qhp->wq);
762 spin_unlock_irqrestore(&qhp->lock, flag);
771 wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
772 qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
775 if (wr->send_flags & IB_SEND_SOLICITED)
776 fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
777 if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
778 fw_flags |= FW_RI_COMPLETION_FLAG;
779 swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
780 switch (wr->opcode) {
781 case IB_WR_SEND_WITH_INV:
783 if (wr->send_flags & IB_SEND_FENCE)
784 fw_flags |= FW_RI_READ_FENCE_FLAG;
785 fw_opcode = FW_RI_SEND_WR;
786 if (wr->opcode == IB_WR_SEND)
787 swsqe->opcode = FW_RI_SEND;
789 swsqe->opcode = FW_RI_SEND_WITH_INV;
790 err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
792 case IB_WR_RDMA_WRITE:
793 fw_opcode = FW_RI_RDMA_WRITE_WR;
794 swsqe->opcode = FW_RI_RDMA_WRITE;
795 err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
797 case IB_WR_RDMA_READ:
798 case IB_WR_RDMA_READ_WITH_INV:
799 fw_opcode = FW_RI_RDMA_READ_WR;
800 swsqe->opcode = FW_RI_READ_REQ;
801 if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
802 fw_flags = FW_RI_RDMA_READ_INVALIDATE;
805 err = build_rdma_read(wqe, wr, &len16);
808 swsqe->read_len = wr->sg_list[0].length;
809 if (!qhp->wq.sq.oldest_read)
810 qhp->wq.sq.oldest_read = swsqe;
813 fw_opcode = FW_RI_FR_NSMR_WR;
814 swsqe->opcode = FW_RI_FAST_REGISTER;
815 err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr), &len16,
817 qhp->rhp->rdev.lldi.adapter_type) ?
820 case IB_WR_LOCAL_INV:
821 if (wr->send_flags & IB_SEND_FENCE)
822 fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
823 fw_opcode = FW_RI_INV_LSTAG_WR;
824 swsqe->opcode = FW_RI_LOCAL_INV;
825 err = build_inv_stag(wqe, wr, &len16);
828 PDBG("%s post of type=%d TBD!\n", __func__,
836 swsqe->idx = qhp->wq.sq.pidx;
838 swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
841 swsqe->wr_id = wr->wr_id;
843 swsqe->sge_ts = cxgb4_read_sge_timestamp(
844 qhp->rhp->rdev.lldi.ports[0]);
845 getnstimeofday(&swsqe->host_ts);
848 init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
850 PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
851 __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
852 swsqe->opcode, swsqe->read_len);
855 t4_sq_produce(&qhp->wq, len16);
856 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
858 if (!qhp->rhp->rdev.status_page->db_off) {
859 t4_ring_sq_db(&qhp->wq, idx, wqe);
860 spin_unlock_irqrestore(&qhp->lock, flag);
862 spin_unlock_irqrestore(&qhp->lock, flag);
863 ring_kernel_sq_db(qhp, idx);
868 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
869 struct ib_recv_wr **bad_wr)
873 union t4_recv_wr *wqe = NULL;
879 qhp = to_c4iw_qp(ibqp);
880 spin_lock_irqsave(&qhp->lock, flag);
881 if (t4_wq_in_error(&qhp->wq)) {
882 spin_unlock_irqrestore(&qhp->lock, flag);
885 num_wrs = t4_rq_avail(&qhp->wq);
887 spin_unlock_irqrestore(&qhp->lock, flag);
891 if (wr->num_sge > T4_MAX_RECV_SGE) {
896 wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
900 err = build_rdma_recv(qhp, wqe, wr, &len16);
908 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
910 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
911 cxgb4_read_sge_timestamp(
912 qhp->rhp->rdev.lldi.ports[0]);
914 &qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_ts);
917 wqe->recv.opcode = FW_RI_RECV_WR;
919 wqe->recv.wrid = qhp->wq.rq.pidx;
923 wqe->recv.len16 = len16;
924 PDBG("%s cookie 0x%llx pidx %u\n", __func__,
925 (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
926 t4_rq_produce(&qhp->wq, len16);
927 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
931 if (!qhp->rhp->rdev.status_page->db_off) {
932 t4_ring_rq_db(&qhp->wq, idx, wqe);
933 spin_unlock_irqrestore(&qhp->lock, flag);
935 spin_unlock_irqrestore(&qhp->lock, flag);
936 ring_kernel_rq_db(qhp, idx);
941 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
946 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
956 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
961 status = CQE_STATUS(err_cqe);
962 opcode = CQE_OPCODE(err_cqe);
963 rqtype = RQ_TYPE(err_cqe);
964 send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
965 (opcode == FW_RI_SEND_WITH_SE_INV);
966 tagged = (opcode == FW_RI_RDMA_WRITE) ||
967 (rqtype && (opcode == FW_RI_READ_RESP));
972 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
973 *ecode = RDMAP_CANT_INV_STAG;
975 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
976 *ecode = RDMAP_INV_STAG;
980 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
981 if ((opcode == FW_RI_SEND_WITH_INV) ||
982 (opcode == FW_RI_SEND_WITH_SE_INV))
983 *ecode = RDMAP_CANT_INV_STAG;
985 *ecode = RDMAP_STAG_NOT_ASSOC;
988 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
989 *ecode = RDMAP_STAG_NOT_ASSOC;
992 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
993 *ecode = RDMAP_ACC_VIOL;
996 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
997 *ecode = RDMAP_TO_WRAP;
1001 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1002 *ecode = DDPT_BASE_BOUNDS;
1004 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1005 *ecode = RDMAP_BASE_BOUNDS;
1008 case T4_ERR_INVALIDATE_SHARED_MR:
1009 case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1010 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1011 *ecode = RDMAP_CANT_INV_STAG;
1014 case T4_ERR_ECC_PSTAG:
1015 case T4_ERR_INTERNAL_ERR:
1016 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1019 case T4_ERR_OUT_OF_RQE:
1020 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1021 *ecode = DDPU_INV_MSN_NOBUF;
1023 case T4_ERR_PBL_ADDR_BOUND:
1024 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1025 *ecode = DDPT_BASE_BOUNDS;
1028 *layer_type = LAYER_MPA|DDP_LLP;
1029 *ecode = MPA_CRC_ERR;
1032 *layer_type = LAYER_MPA|DDP_LLP;
1033 *ecode = MPA_MARKER_ERR;
1035 case T4_ERR_PDU_LEN_ERR:
1036 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1037 *ecode = DDPU_MSG_TOOBIG;
1039 case T4_ERR_DDP_VERSION:
1041 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1042 *ecode = DDPT_INV_VERS;
1044 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1045 *ecode = DDPU_INV_VERS;
1048 case T4_ERR_RDMA_VERSION:
1049 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1050 *ecode = RDMAP_INV_VERS;
1053 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1054 *ecode = RDMAP_INV_OPCODE;
1056 case T4_ERR_DDP_QUEUE_NUM:
1057 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1058 *ecode = DDPU_INV_QN;
1061 case T4_ERR_MSN_GAP:
1062 case T4_ERR_MSN_RANGE:
1063 case T4_ERR_IRD_OVERFLOW:
1064 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1065 *ecode = DDPU_INV_MSN_RANGE;
1068 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1072 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1073 *ecode = DDPU_INV_MO;
1076 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1082 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1085 struct fw_ri_wr *wqe;
1086 struct sk_buff *skb;
1087 struct terminate_message *term;
1089 PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1092 skb = alloc_skb(sizeof *wqe, gfp);
1095 set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1097 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1098 memset(wqe, 0, sizeof *wqe);
1099 wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1100 wqe->flowid_len16 = cpu_to_be32(
1101 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1102 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1104 wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1105 wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1106 term = (struct terminate_message *)wqe->u.terminate.termmsg;
1107 if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1108 term->layer_etype = qhp->attr.layer_etype;
1109 term->ecode = qhp->attr.ecode;
1111 build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1112 c4iw_ofld_send(&qhp->rhp->rdev, skb);
1116 * Assumes qhp lock is held.
1118 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1119 struct c4iw_cq *schp)
1122 int rq_flushed, sq_flushed;
1125 PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1127 /* locking hierarchy: cq lock first, then qp lock. */
1128 spin_lock_irqsave(&rchp->lock, flag);
1129 spin_lock(&qhp->lock);
1131 if (qhp->wq.flushed) {
1132 spin_unlock(&qhp->lock);
1133 spin_unlock_irqrestore(&rchp->lock, flag);
1136 qhp->wq.flushed = 1;
1138 c4iw_flush_hw_cq(rchp);
1139 c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1140 rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1141 spin_unlock(&qhp->lock);
1142 spin_unlock_irqrestore(&rchp->lock, flag);
1144 /* locking hierarchy: cq lock first, then qp lock. */
1145 spin_lock_irqsave(&schp->lock, flag);
1146 spin_lock(&qhp->lock);
1148 c4iw_flush_hw_cq(schp);
1149 sq_flushed = c4iw_flush_sq(qhp);
1150 spin_unlock(&qhp->lock);
1151 spin_unlock_irqrestore(&schp->lock, flag);
1154 if (t4_clear_cq_armed(&rchp->cq) &&
1155 (rq_flushed || sq_flushed)) {
1156 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1157 (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1158 rchp->ibcq.cq_context);
1159 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1162 if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1163 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1164 (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1165 rchp->ibcq.cq_context);
1166 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1168 if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1169 spin_lock_irqsave(&schp->comp_handler_lock, flag);
1170 (*schp->ibcq.comp_handler)(&schp->ibcq,
1171 schp->ibcq.cq_context);
1172 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1177 static void flush_qp(struct c4iw_qp *qhp)
1179 struct c4iw_cq *rchp, *schp;
1182 rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1183 schp = to_c4iw_cq(qhp->ibqp.send_cq);
1185 t4_set_wq_in_error(&qhp->wq);
1186 if (qhp->ibqp.uobject) {
1188 /* for user qps, qhp->wq.flushed is protected by qhp->mutex */
1189 if (qhp->wq.flushed)
1192 qhp->wq.flushed = 1;
1193 t4_set_cq_in_error(&rchp->cq);
1194 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1195 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1196 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1198 t4_set_cq_in_error(&schp->cq);
1199 spin_lock_irqsave(&schp->comp_handler_lock, flag);
1200 (*schp->ibcq.comp_handler)(&schp->ibcq,
1201 schp->ibcq.cq_context);
1202 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1206 __flush_qp(qhp, rchp, schp);
1209 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1212 struct fw_ri_wr *wqe;
1214 struct sk_buff *skb;
1216 PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1219 skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1222 set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1224 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1225 memset(wqe, 0, sizeof *wqe);
1226 wqe->op_compl = cpu_to_be32(
1227 FW_WR_OP_V(FW_RI_INIT_WR) |
1229 wqe->flowid_len16 = cpu_to_be32(
1230 FW_WR_FLOWID_V(ep->hwtid) |
1231 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1232 wqe->cookie = (uintptr_t)&ep->com.wr_wait;
1234 wqe->u.fini.type = FW_RI_TYPE_FINI;
1235 ret = c4iw_ofld_send(&rhp->rdev, skb);
1239 ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1240 qhp->wq.sq.qid, __func__);
1242 PDBG("%s ret %d\n", __func__, ret);
1246 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1248 PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1249 memset(&init->u, 0, sizeof init->u);
1251 case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1252 init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1253 init->u.write.stag_sink = cpu_to_be32(1);
1254 init->u.write.to_sink = cpu_to_be64(1);
1255 init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1256 init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1257 sizeof(struct fw_ri_immd),
1260 case FW_RI_INIT_P2PTYPE_READ_REQ:
1261 init->u.write.opcode = FW_RI_RDMA_READ_WR;
1262 init->u.read.stag_src = cpu_to_be32(1);
1263 init->u.read.to_src_lo = cpu_to_be32(1);
1264 init->u.read.stag_sink = cpu_to_be32(1);
1265 init->u.read.to_sink_lo = cpu_to_be32(1);
1266 init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1271 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1273 struct fw_ri_wr *wqe;
1275 struct sk_buff *skb;
1277 PDBG("%s qhp %p qid 0x%x tid %u ird %u ord %u\n", __func__, qhp,
1278 qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1280 skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1285 ret = alloc_ird(rhp, qhp->attr.max_ird);
1287 qhp->attr.max_ird = 0;
1291 set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1293 wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1294 memset(wqe, 0, sizeof *wqe);
1295 wqe->op_compl = cpu_to_be32(
1296 FW_WR_OP_V(FW_RI_INIT_WR) |
1298 wqe->flowid_len16 = cpu_to_be32(
1299 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1300 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1302 wqe->cookie = (uintptr_t)&qhp->ep->com.wr_wait;
1304 wqe->u.init.type = FW_RI_TYPE_INIT;
1305 wqe->u.init.mpareqbit_p2ptype =
1306 FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1307 FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1308 wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1309 if (qhp->attr.mpa_attr.recv_marker_enabled)
1310 wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1311 if (qhp->attr.mpa_attr.xmit_marker_enabled)
1312 wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1313 if (qhp->attr.mpa_attr.crc_enabled)
1314 wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1316 wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1317 FW_RI_QP_RDMA_WRITE_ENABLE |
1318 FW_RI_QP_BIND_ENABLE;
1319 if (!qhp->ibqp.uobject)
1320 wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1321 FW_RI_QP_STAG0_ENABLE;
1322 wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1323 wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1324 wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1325 wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1326 wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1327 wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1328 wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1329 wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1330 wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1331 wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1332 wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1333 wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1334 wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1335 rhp->rdev.lldi.vr->rq.start);
1336 if (qhp->attr.mpa_attr.initiator)
1337 build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1339 ret = c4iw_ofld_send(&rhp->rdev, skb);
1343 ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1344 qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1348 free_ird(rhp, qhp->attr.max_ird);
1350 PDBG("%s ret %d\n", __func__, ret);
1354 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1355 enum c4iw_qp_attr_mask mask,
1356 struct c4iw_qp_attributes *attrs,
1360 struct c4iw_qp_attributes newattr = qhp->attr;
1365 struct c4iw_ep *ep = NULL;
1367 PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1368 qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1369 (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1371 mutex_lock(&qhp->mutex);
1373 /* Process attr changes if in IDLE */
1374 if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1375 if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1379 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1380 newattr.enable_rdma_read = attrs->enable_rdma_read;
1381 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1382 newattr.enable_rdma_write = attrs->enable_rdma_write;
1383 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1384 newattr.enable_bind = attrs->enable_bind;
1385 if (mask & C4IW_QP_ATTR_MAX_ORD) {
1386 if (attrs->max_ord > c4iw_max_read_depth) {
1390 newattr.max_ord = attrs->max_ord;
1392 if (mask & C4IW_QP_ATTR_MAX_IRD) {
1393 if (attrs->max_ird > cur_max_read_depth(rhp)) {
1397 newattr.max_ird = attrs->max_ird;
1399 qhp->attr = newattr;
1402 if (mask & C4IW_QP_ATTR_SQ_DB) {
1403 ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1406 if (mask & C4IW_QP_ATTR_RQ_DB) {
1407 ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1411 if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1413 if (qhp->attr.state == attrs->next_state)
1416 switch (qhp->attr.state) {
1417 case C4IW_QP_STATE_IDLE:
1418 switch (attrs->next_state) {
1419 case C4IW_QP_STATE_RTS:
1420 if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1424 if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1428 qhp->attr.mpa_attr = attrs->mpa_attr;
1429 qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1430 qhp->ep = qhp->attr.llp_stream_handle;
1431 set_state(qhp, C4IW_QP_STATE_RTS);
1434 * Ref the endpoint here and deref when we
1435 * disassociate the endpoint from the QP. This
1436 * happens in CLOSING->IDLE transition or *->ERROR
1439 c4iw_get_ep(&qhp->ep->com);
1440 ret = rdma_init(rhp, qhp);
1444 case C4IW_QP_STATE_ERROR:
1445 set_state(qhp, C4IW_QP_STATE_ERROR);
1453 case C4IW_QP_STATE_RTS:
1454 switch (attrs->next_state) {
1455 case C4IW_QP_STATE_CLOSING:
1456 BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1457 t4_set_wq_in_error(&qhp->wq);
1458 set_state(qhp, C4IW_QP_STATE_CLOSING);
1463 c4iw_get_ep(&qhp->ep->com);
1465 ret = rdma_fini(rhp, qhp, ep);
1469 case C4IW_QP_STATE_TERMINATE:
1470 t4_set_wq_in_error(&qhp->wq);
1471 set_state(qhp, C4IW_QP_STATE_TERMINATE);
1472 qhp->attr.layer_etype = attrs->layer_etype;
1473 qhp->attr.ecode = attrs->ecode;
1476 c4iw_get_ep(&qhp->ep->com);
1480 terminate = qhp->attr.send_term;
1481 ret = rdma_fini(rhp, qhp, ep);
1486 case C4IW_QP_STATE_ERROR:
1487 t4_set_wq_in_error(&qhp->wq);
1488 set_state(qhp, C4IW_QP_STATE_ERROR);
1493 c4iw_get_ep(&qhp->ep->com);
1502 case C4IW_QP_STATE_CLOSING:
1507 switch (attrs->next_state) {
1508 case C4IW_QP_STATE_IDLE:
1510 set_state(qhp, C4IW_QP_STATE_IDLE);
1511 qhp->attr.llp_stream_handle = NULL;
1512 c4iw_put_ep(&qhp->ep->com);
1514 wake_up(&qhp->wait);
1516 case C4IW_QP_STATE_ERROR:
1523 case C4IW_QP_STATE_ERROR:
1524 if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1528 if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1532 set_state(qhp, C4IW_QP_STATE_IDLE);
1534 case C4IW_QP_STATE_TERMINATE:
1542 printk(KERN_ERR "%s in a bad state %d\n",
1543 __func__, qhp->attr.state);
1550 PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1553 /* disassociate the LLP connection */
1554 qhp->attr.llp_stream_handle = NULL;
1558 set_state(qhp, C4IW_QP_STATE_ERROR);
1563 wake_up(&qhp->wait);
1565 mutex_unlock(&qhp->mutex);
1568 post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1571 * If disconnect is 1, then we need to initiate a disconnect
1572 * on the EP. This can be a normal close (RTS->CLOSING) or
1573 * an abnormal close (RTS/CLOSING->ERROR).
1576 c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1578 c4iw_put_ep(&ep->com);
1582 * If free is 1, then we've disassociated the EP from the QP
1583 * and we need to dereference the EP.
1586 c4iw_put_ep(&ep->com);
1587 PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1591 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1593 struct c4iw_dev *rhp;
1594 struct c4iw_qp *qhp;
1595 struct c4iw_qp_attributes attrs;
1596 struct c4iw_ucontext *ucontext;
1598 qhp = to_c4iw_qp(ib_qp);
1601 attrs.next_state = C4IW_QP_STATE_ERROR;
1602 if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1603 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1605 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1606 wait_event(qhp->wait, !qhp->ep);
1608 remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1609 atomic_dec(&qhp->refcnt);
1610 wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1612 spin_lock_irq(&rhp->lock);
1613 if (!list_empty(&qhp->db_fc_entry))
1614 list_del_init(&qhp->db_fc_entry);
1615 spin_unlock_irq(&rhp->lock);
1616 free_ird(rhp, qhp->attr.max_ird);
1618 ucontext = ib_qp->uobject ?
1619 to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1620 destroy_qp(&rhp->rdev, &qhp->wq,
1621 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1623 PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1628 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1629 struct ib_udata *udata)
1631 struct c4iw_dev *rhp;
1632 struct c4iw_qp *qhp;
1633 struct c4iw_pd *php;
1634 struct c4iw_cq *schp;
1635 struct c4iw_cq *rchp;
1636 struct c4iw_create_qp_resp uresp;
1637 unsigned int sqsize, rqsize;
1638 struct c4iw_ucontext *ucontext;
1640 struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1642 PDBG("%s ib_pd %p\n", __func__, pd);
1644 if (attrs->qp_type != IB_QPT_RC)
1645 return ERR_PTR(-EINVAL);
1647 php = to_c4iw_pd(pd);
1649 schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1650 rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1652 return ERR_PTR(-EINVAL);
1654 if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1655 return ERR_PTR(-EINVAL);
1657 if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1658 return ERR_PTR(-E2BIG);
1659 rqsize = attrs->cap.max_recv_wr + 1;
1663 if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1664 return ERR_PTR(-E2BIG);
1665 sqsize = attrs->cap.max_send_wr + 1;
1669 ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1671 qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1673 return ERR_PTR(-ENOMEM);
1674 qhp->wq.sq.size = sqsize;
1675 qhp->wq.sq.memsize =
1676 (sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1677 sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1678 qhp->wq.sq.flush_cidx = -1;
1679 qhp->wq.rq.size = rqsize;
1680 qhp->wq.rq.memsize =
1681 (rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1682 sizeof(*qhp->wq.rq.queue);
1685 qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1686 qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1689 ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1690 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1694 attrs->cap.max_recv_wr = rqsize - 1;
1695 attrs->cap.max_send_wr = sqsize - 1;
1696 attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1699 qhp->attr.pd = php->pdid;
1700 qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1701 qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1702 qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1703 qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1704 qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1705 qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1706 qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1707 qhp->attr.state = C4IW_QP_STATE_IDLE;
1708 qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1709 qhp->attr.enable_rdma_read = 1;
1710 qhp->attr.enable_rdma_write = 1;
1711 qhp->attr.enable_bind = 1;
1712 qhp->attr.max_ord = 0;
1713 qhp->attr.max_ird = 0;
1714 qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1715 spin_lock_init(&qhp->lock);
1716 mutex_init(&qhp->mutex);
1717 init_waitqueue_head(&qhp->wait);
1718 atomic_set(&qhp->refcnt, 1);
1720 ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1725 mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1730 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1735 mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1740 mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1745 if (t4_sq_onchip(&qhp->wq.sq)) {
1746 mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1751 uresp.flags = C4IW_QPF_ONCHIP;
1754 uresp.qid_mask = rhp->rdev.qpmask;
1755 uresp.sqid = qhp->wq.sq.qid;
1756 uresp.sq_size = qhp->wq.sq.size;
1757 uresp.sq_memsize = qhp->wq.sq.memsize;
1758 uresp.rqid = qhp->wq.rq.qid;
1759 uresp.rq_size = qhp->wq.rq.size;
1760 uresp.rq_memsize = qhp->wq.rq.memsize;
1761 spin_lock(&ucontext->mmap_lock);
1763 uresp.ma_sync_key = ucontext->key;
1764 ucontext->key += PAGE_SIZE;
1766 uresp.ma_sync_key = 0;
1768 uresp.sq_key = ucontext->key;
1769 ucontext->key += PAGE_SIZE;
1770 uresp.rq_key = ucontext->key;
1771 ucontext->key += PAGE_SIZE;
1772 uresp.sq_db_gts_key = ucontext->key;
1773 ucontext->key += PAGE_SIZE;
1774 uresp.rq_db_gts_key = ucontext->key;
1775 ucontext->key += PAGE_SIZE;
1776 spin_unlock(&ucontext->mmap_lock);
1777 ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1780 mm1->key = uresp.sq_key;
1781 mm1->addr = qhp->wq.sq.phys_addr;
1782 mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1783 insert_mmap(ucontext, mm1);
1784 mm2->key = uresp.rq_key;
1785 mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1786 mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1787 insert_mmap(ucontext, mm2);
1788 mm3->key = uresp.sq_db_gts_key;
1789 mm3->addr = (__force unsigned long)qhp->wq.sq.bar2_pa;
1790 mm3->len = PAGE_SIZE;
1791 insert_mmap(ucontext, mm3);
1792 mm4->key = uresp.rq_db_gts_key;
1793 mm4->addr = (__force unsigned long)qhp->wq.rq.bar2_pa;
1794 mm4->len = PAGE_SIZE;
1795 insert_mmap(ucontext, mm4);
1797 mm5->key = uresp.ma_sync_key;
1798 mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1799 + PCIE_MA_SYNC_A) & PAGE_MASK;
1800 mm5->len = PAGE_SIZE;
1801 insert_mmap(ucontext, mm5);
1804 qhp->ibqp.qp_num = qhp->wq.sq.qid;
1805 init_timer(&(qhp->timer));
1806 INIT_LIST_HEAD(&qhp->db_fc_entry);
1807 PDBG("%s sq id %u size %u memsize %zu num_entries %u "
1808 "rq id %u size %u memsize %zu num_entries %u\n", __func__,
1809 qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
1810 attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
1811 qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1824 remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1826 destroy_qp(&rhp->rdev, &qhp->wq,
1827 ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1830 return ERR_PTR(ret);
1833 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1834 int attr_mask, struct ib_udata *udata)
1836 struct c4iw_dev *rhp;
1837 struct c4iw_qp *qhp;
1838 enum c4iw_qp_attr_mask mask = 0;
1839 struct c4iw_qp_attributes attrs;
1841 PDBG("%s ib_qp %p\n", __func__, ibqp);
1843 /* iwarp does not support the RTR state */
1844 if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1845 attr_mask &= ~IB_QP_STATE;
1847 /* Make sure we still have something left to do */
1851 memset(&attrs, 0, sizeof attrs);
1852 qhp = to_c4iw_qp(ibqp);
1855 attrs.next_state = c4iw_convert_state(attr->qp_state);
1856 attrs.enable_rdma_read = (attr->qp_access_flags &
1857 IB_ACCESS_REMOTE_READ) ? 1 : 0;
1858 attrs.enable_rdma_write = (attr->qp_access_flags &
1859 IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1860 attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1863 mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1864 mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1865 (C4IW_QP_ATTR_ENABLE_RDMA_READ |
1866 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1867 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1870 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1871 * ringing the queue db when we're in DB_FULL mode.
1872 * Only allow this on T4 devices.
1874 attrs.sq_db_inc = attr->sq_psn;
1875 attrs.rq_db_inc = attr->rq_psn;
1876 mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1877 mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1878 if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
1879 (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
1882 return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1885 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1887 PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1888 return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1891 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1892 int attr_mask, struct ib_qp_init_attr *init_attr)
1894 struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1896 memset(attr, 0, sizeof *attr);
1897 memset(init_attr, 0, sizeof *init_attr);
1898 attr->qp_state = to_ib_qp_state(qhp->attr.state);
1899 init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1900 init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1901 init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1902 init_attr->cap.max_recv_sge = qhp->attr.rq_max_sges;
1903 init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1904 init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;