xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/qp.c (revision 05057380ca392b29d33a4d362936e8af81101685)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2009-2013 Chelsio, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 #include <sys/cdefs.h>
35 #include "opt_inet.h"
36 
37 #ifdef TCP_OFFLOAD
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>
46 
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>
54 
55 #include <netinet/toecore.h>
56 
57 struct sge_iq;
58 struct rss_header;
59 struct cpl_set_tcb_rpl;
60 #include <linux/types.h>
61 #include "offload.h"
62 #include "tom/t4_tom.h"
63 
64 #include "iw_cxgbe.h"
65 #include "user.h"
66 
67 static int creds(struct toepcb *toep, struct tcpcb *tp, size_t wrsize);
68 static int max_fr_immd = T4_MAX_FR_IMMD;//SYSCTL parameter later...
69 
alloc_ird(struct c4iw_dev * dev,u32 ird)70 static int alloc_ird(struct c4iw_dev *dev, u32 ird)
71 {
72 	int ret = 0;
73 
74 	xa_lock_irq(&dev->qps);
75 	if (ird <= dev->avail_ird)
76 		dev->avail_ird -= ird;
77 	else
78 		ret = -ENOMEM;
79 	xa_unlock_irq(&dev->qps);
80 
81 	if (ret)
82 		log(LOG_WARNING, "%s: device IRD resources exhausted\n",
83 			device_get_nameunit(dev->rdev.adap->dev));
84 
85 	return ret;
86 }
87 
free_ird(struct c4iw_dev * dev,int ird)88 static void free_ird(struct c4iw_dev *dev, int ird)
89 {
90 	xa_lock_irq(&dev->qps);
91 	dev->avail_ird += ird;
92 	xa_unlock_irq(&dev->qps);
93 }
94 
set_state(struct c4iw_qp * qhp,enum c4iw_qp_state state)95 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
96 {
97 	unsigned long flag;
98 	spin_lock_irqsave(&qhp->lock, flag);
99 	qhp->attr.state = state;
100 	spin_unlock_irqrestore(&qhp->lock, flag);
101 }
102 
destroy_qp(struct c4iw_rdev * rdev,struct t4_wq * wq,struct c4iw_dev_ucontext * uctx)103 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
104 		      struct c4iw_dev_ucontext *uctx)
105 {
106 	struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev);
107 	/*
108 	 * uP clears EQ contexts when the connection exits rdma mode,
109 	 * so no need to post a RESET WR for these EQs.
110 	 */
111 	dma_free_coherent(rhp->ibdev.dma_device,
112 			wq->rq.memsize, wq->rq.queue,
113 			dma_unmap_addr(&wq->rq, mapping));
114 	dma_free_coherent(rhp->ibdev.dma_device,
115 			wq->sq.memsize, wq->sq.queue,
116 			dma_unmap_addr(&wq->sq, mapping));
117 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
118 	kfree(wq->rq.sw_rq);
119 	kfree(wq->sq.sw_sq);
120 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
121 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
122 	return 0;
123 }
124 
create_qp(struct c4iw_rdev * rdev,struct t4_wq * wq,struct t4_cq * rcq,struct t4_cq * scq,struct c4iw_dev_ucontext * uctx,struct c4iw_wr_wait * wr_waitp)125 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
126 		     struct t4_cq *rcq, struct t4_cq *scq,
127 		     struct c4iw_dev_ucontext *uctx,
128 		     struct c4iw_wr_wait *wr_waitp)
129 {
130 	struct adapter *sc = rdev->adap;
131 	struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev);
132 	int user = (uctx != &rdev->uctx);
133 	struct fw_ri_res_wr *res_wr;
134 	struct fw_ri_res *res;
135 	int wr_len;
136 	int ret = 0;
137 	int eqsize;
138 	struct wrqe *wr;
139 	u64 sq_bar2_qoffset = 0, rq_bar2_qoffset = 0;
140 
141 	if (__predict_false(c4iw_stopped(rdev)))
142 		return -EIO;
143 
144 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
145 	if (!wq->sq.qid)
146 		return -ENOMEM;
147 
148 	wq->rq.qid = c4iw_get_qpid(rdev, uctx);
149 	if (!wq->rq.qid) {
150 		ret = -ENOMEM;
151 		goto free_sq_qid;
152 	}
153 
154 	if (!user) {
155 		wq->sq.sw_sq = kcalloc(wq->sq.size, sizeof(*wq->sq.sw_sq),
156 				       GFP_KERNEL);
157 		if (!wq->sq.sw_sq) {
158 			ret = -ENOMEM;
159 			goto free_rq_qid;
160 		}
161 
162 		wq->rq.sw_rq = kcalloc(wq->rq.size, sizeof(*wq->rq.sw_rq),
163 				       GFP_KERNEL);
164 		if (!wq->rq.sw_rq) {
165 			ret = -ENOMEM;
166 			goto free_sw_sq;
167 		}
168 	}
169 
170 	/*
171 	 * RQT must be a power of 2 and at least 16 deep.
172 	 */
173 	wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
174 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
175 	if (!wq->rq.rqt_hwaddr) {
176 		ret = -ENOMEM;
177 		goto free_sw_rq;
178 	}
179 
180 	/*QP memory, allocate DMAable memory for Send & Receive Queues */
181 	wq->sq.queue = dma_alloc_coherent(rhp->ibdev.dma_device, wq->sq.memsize,
182 				       &(wq->sq.dma_addr), GFP_KERNEL);
183 	if (!wq->sq.queue) {
184 		ret = -ENOMEM;
185 		goto free_hwaddr;
186 	}
187 	wq->sq.phys_addr = vtophys(wq->sq.queue);
188 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
189 	memset(wq->sq.queue, 0, wq->sq.memsize);
190 
191 	wq->rq.queue = dma_alloc_coherent(rhp->ibdev.dma_device,
192 			wq->rq.memsize, &(wq->rq.dma_addr), GFP_KERNEL);
193 	if (!wq->rq.queue) {
194 		ret = -ENOMEM;
195 		goto free_sq_dma;
196 	}
197 	wq->rq.phys_addr = vtophys(wq->rq.queue);
198 	dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
199 
200 	CTR5(KTR_IW_CXGBE,
201 	    "%s QP sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx",
202 	    __func__,
203 	    wq->sq.queue, (unsigned long long)wq->sq.phys_addr,
204 	    wq->rq.queue, (unsigned long long)wq->rq.phys_addr);
205 
206 	/* Doorbell/WC regions, determine the BAR2 queue offset and qid. */
207 	t4_bar2_sge_qregs(rdev->adap, wq->sq.qid, T4_BAR2_QTYPE_EGRESS, user,
208 			&sq_bar2_qoffset, &wq->sq.bar2_qid);
209 	t4_bar2_sge_qregs(rdev->adap, wq->rq.qid, T4_BAR2_QTYPE_EGRESS, user,
210 			&rq_bar2_qoffset, &wq->rq.bar2_qid);
211 
212 	if (user) {
213 		/* Compute BAR2 DB/WC physical address(page-aligned) for
214 		 * Userspace mapping.
215 		 */
216 		wq->sq.bar2_pa = (rdev->bar2_pa + sq_bar2_qoffset) & PAGE_MASK;
217 		wq->rq.bar2_pa = (rdev->bar2_pa + rq_bar2_qoffset) & PAGE_MASK;
218 		CTR3(KTR_IW_CXGBE,
219 			"%s BAR2 DB/WC sq base pa 0x%llx rq base pa 0x%llx",
220 			__func__, (unsigned long long)wq->sq.bar2_pa,
221 			(unsigned long long)wq->rq.bar2_pa);
222 	} else {
223 		/* Compute BAR2 DB/WC virtual address to access in kernel. */
224 		wq->sq.bar2_va = (void __iomem *)((u64)rdev->bar2_kva +
225 				sq_bar2_qoffset);
226 		wq->rq.bar2_va = (void __iomem *)((u64)rdev->bar2_kva +
227 				rq_bar2_qoffset);
228 		CTR3(KTR_IW_CXGBE, "%s BAR2 DB/WC sq base va %p rq base va %p",
229 			__func__, (unsigned long long)wq->sq.bar2_va,
230 			(unsigned long long)wq->rq.bar2_va);
231 	}
232 
233 	wq->rdev = rdev;
234 	wq->rq.msn = 1;
235 
236 	/* build fw_ri_res_wr */
237 	wr_len = sizeof *res_wr + 2 * sizeof *res;
238 
239 	wr = alloc_wrqe(wr_len, &sc->sge.ctrlq[0]);
240 	if (wr == NULL) {
241 		ret = -ENOMEM;
242 		goto free_rq_dma;
243 	}
244         res_wr = wrtod(wr);
245 
246 	memset(res_wr, 0, wr_len);
247 	res_wr->op_nres = cpu_to_be32(
248 			V_FW_WR_OP(FW_RI_RES_WR) |
249 			V_FW_RI_RES_WR_NRES(2) |
250 			F_FW_WR_COMPL);
251 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
252 	res_wr->cookie = (uintptr_t)wr_waitp;
253 	res = res_wr->res;
254 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
255 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
256 
257 	/* eqsize is the number of 64B entries plus the status page size. */
258 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
259 			rdev->hw_queue.t4_eq_status_entries;
260 
261 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
262 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
263 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
264 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
265 		V_FW_RI_RES_WR_IQID(scq->cqid));
266 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
267 		V_FW_RI_RES_WR_DCAEN(0) |
268 		V_FW_RI_RES_WR_DCACPU(0) |
269 		V_FW_RI_RES_WR_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
270 		    X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
271 		V_FW_RI_RES_WR_FBMAX(3) |
272 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
273 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
274 		V_FW_RI_RES_WR_EQSIZE(eqsize));
275 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
276 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
277 	res++;
278 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
279 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
280 
281 	/* eqsize is the number of 64B entries plus the status page size. */
282 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
283 			rdev->hw_queue.t4_eq_status_entries;
284 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
285 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
286 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
287 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
288 		V_FW_RI_RES_WR_IQID(rcq->cqid));
289 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
290 		V_FW_RI_RES_WR_DCAEN(0) |
291 		V_FW_RI_RES_WR_DCACPU(0) |
292 		V_FW_RI_RES_WR_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
293 		    X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
294 		V_FW_RI_RES_WR_FBMAX(3) |
295 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
296 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
297 		V_FW_RI_RES_WR_EQSIZE(eqsize));
298 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
299 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
300 
301 	c4iw_init_wr_wait(wr_waitp);
302 
303 	ret = c4iw_ref_send_wait(rdev, wr, wr_waitp, 0, wq->sq.qid,
304 			NULL, __func__);
305 	if (ret)
306 		goto free_rq_dma;
307 
308 	CTR5(KTR_IW_CXGBE,
309 	    "%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx",
310 	    __func__, wq->sq.qid, wq->rq.qid,
311 	    (unsigned long long)wq->sq.bar2_va,
312 	    (unsigned long long)wq->rq.bar2_va);
313 
314 	return 0;
315 free_rq_dma:
316 	dma_free_coherent(rhp->ibdev.dma_device,
317 			  wq->rq.memsize, wq->rq.queue,
318 			  dma_unmap_addr(&wq->rq, mapping));
319 free_sq_dma:
320 	dma_free_coherent(rhp->ibdev.dma_device,
321 			  wq->sq.memsize, wq->sq.queue,
322 			  dma_unmap_addr(&wq->sq, mapping));
323 free_hwaddr:
324 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
325 free_sw_rq:
326 	kfree(wq->rq.sw_rq);
327 free_sw_sq:
328 	kfree(wq->sq.sw_sq);
329 free_rq_qid:
330 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
331 free_sq_qid:
332 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
333 	return ret;
334 }
335 
build_immd(struct t4_sq * sq,struct fw_ri_immd * immdp,const struct ib_send_wr * wr,int max,u32 * plenp)336 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
337 		      const struct ib_send_wr *wr, int max, u32 *plenp)
338 {
339 	u8 *dstp, *srcp;
340 	u32 plen = 0;
341 	int i;
342 	int rem, len;
343 
344 	dstp = (u8 *)immdp->data;
345 	for (i = 0; i < wr->num_sge; i++) {
346 		if ((plen + wr->sg_list[i].length) > max)
347 			return -EMSGSIZE;
348 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
349 		plen += wr->sg_list[i].length;
350 		rem = wr->sg_list[i].length;
351 		while (rem) {
352 			if (dstp == (u8 *)&sq->queue[sq->size])
353 				dstp = (u8 *)sq->queue;
354 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
355 				len = rem;
356 			else
357 				len = (u8 *)&sq->queue[sq->size] - dstp;
358 			memcpy(dstp, srcp, len);
359 			dstp += len;
360 			srcp += len;
361 			rem -= len;
362 		}
363 	}
364 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
365 	if (len)
366 		memset(dstp, 0, len);
367 	immdp->op = FW_RI_DATA_IMMD;
368 	immdp->r1 = 0;
369 	immdp->r2 = 0;
370 	immdp->immdlen = cpu_to_be32(plen);
371 	*plenp = plen;
372 	return 0;
373 }
374 
build_isgl(__be64 * queue_start,__be64 * queue_end,struct fw_ri_isgl * isglp,struct ib_sge * sg_list,int num_sge,u32 * plenp)375 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
376 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
377 		      int num_sge, u32 *plenp)
378 
379 {
380 	int i;
381 	u32 plen = 0;
382 	__be64 *flitp;
383 
384 	if ((__be64 *)isglp == queue_end)
385 		isglp = (struct fw_ri_isgl *)queue_start;
386 
387 	flitp = (__be64 *)isglp->sge;
388 
389 	for (i = 0; i < num_sge; i++) {
390 		if ((plen + sg_list[i].length) < plen)
391 			return -EMSGSIZE;
392 		plen += sg_list[i].length;
393 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
394 				     sg_list[i].length);
395 		if (++flitp == queue_end)
396 			flitp = queue_start;
397 		*flitp = cpu_to_be64(sg_list[i].addr);
398 		if (++flitp == queue_end)
399 			flitp = queue_start;
400 	}
401 	*flitp = (__force __be64)0;
402 	isglp->op = FW_RI_DATA_ISGL;
403 	isglp->r1 = 0;
404 	isglp->nsge = cpu_to_be16(num_sge);
405 	isglp->r2 = 0;
406 	if (plenp)
407 		*plenp = plen;
408 	return 0;
409 }
410 
build_rdma_send(struct t4_sq * sq,union t4_wr * wqe,const struct ib_send_wr * wr,u8 * len16)411 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
412 			   const struct ib_send_wr *wr, u8 *len16)
413 {
414 	u32 plen;
415 	int size;
416 	int ret;
417 
418 	if (wr->num_sge > T4_MAX_SEND_SGE)
419 		return -EINVAL;
420 	switch (wr->opcode) {
421 	case IB_WR_SEND:
422 		if (wr->send_flags & IB_SEND_SOLICITED)
423 			wqe->send.sendop_pkd = cpu_to_be32(
424 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
425 		else
426 			wqe->send.sendop_pkd = cpu_to_be32(
427 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
428 		wqe->send.stag_inv = 0;
429 		break;
430 	case IB_WR_SEND_WITH_INV:
431 		if (wr->send_flags & IB_SEND_SOLICITED)
432 			wqe->send.sendop_pkd = cpu_to_be32(
433 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
434 		else
435 			wqe->send.sendop_pkd = cpu_to_be32(
436 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
437 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
438 		break;
439 
440 	default:
441 		return -EINVAL;
442 	}
443 	wqe->send.r3 = 0;
444 	wqe->send.r4 = 0;
445 
446 	plen = 0;
447 	if (wr->num_sge) {
448 		if (wr->send_flags & IB_SEND_INLINE) {
449 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
450 					 T4_MAX_SEND_INLINE, &plen);
451 			if (ret)
452 				return ret;
453 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
454 			       plen;
455 		} else {
456 			ret = build_isgl((__be64 *)sq->queue,
457 					 (__be64 *)&sq->queue[sq->size],
458 					 wqe->send.u.isgl_src,
459 					 wr->sg_list, wr->num_sge, &plen);
460 			if (ret)
461 				return ret;
462 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
463 			       wr->num_sge * sizeof(struct fw_ri_sge);
464 		}
465 	} else {
466 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
467 		wqe->send.u.immd_src[0].r1 = 0;
468 		wqe->send.u.immd_src[0].r2 = 0;
469 		wqe->send.u.immd_src[0].immdlen = 0;
470 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
471 		plen = 0;
472 	}
473 	*len16 = DIV_ROUND_UP(size, 16);
474 	wqe->send.plen = cpu_to_be32(plen);
475 	return 0;
476 }
477 
build_rdma_write(struct t4_sq * sq,union t4_wr * wqe,const struct ib_send_wr * wr,u8 * len16)478 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
479 			    const struct ib_send_wr *wr, u8 *len16)
480 {
481 	u32 plen;
482 	int size;
483 	int ret;
484 
485 	if (wr->num_sge > T4_MAX_SEND_SGE)
486 		return -EINVAL;
487 	if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
488 		wqe->write.immd_data = wr->ex.imm_data;
489 	else
490 		wqe->write.immd_data = 0;
491 	wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
492 	wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
493 	if (wr->num_sge) {
494 		if (wr->send_flags & IB_SEND_INLINE) {
495 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
496 					 T4_MAX_WRITE_INLINE, &plen);
497 			if (ret)
498 				return ret;
499 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
500 			       plen;
501 		} else {
502 			ret = build_isgl((__be64 *)sq->queue,
503 					 (__be64 *)&sq->queue[sq->size],
504 					 wqe->write.u.isgl_src,
505 					 wr->sg_list, wr->num_sge, &plen);
506 			if (ret)
507 				return ret;
508 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
509 			       wr->num_sge * sizeof(struct fw_ri_sge);
510 		}
511 	} else {
512 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
513 		wqe->write.u.immd_src[0].r1 = 0;
514 		wqe->write.u.immd_src[0].r2 = 0;
515 		wqe->write.u.immd_src[0].immdlen = 0;
516 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
517 		plen = 0;
518 	}
519 	*len16 = DIV_ROUND_UP(size, 16);
520 	wqe->write.plen = cpu_to_be32(plen);
521 	return 0;
522 }
523 
build_immd_cmpl(struct t4_sq * sq,struct fw_ri_immd_cmpl * immdp,const struct ib_send_wr * wr)524 static void build_immd_cmpl(struct t4_sq *sq, struct fw_ri_immd_cmpl *immdp,
525 			   const struct ib_send_wr *wr)
526 {
527 	memcpy((u8 *)immdp->data, (u8 *)(uintptr_t)wr->sg_list->addr, 16);
528 	memset(immdp->r1, 0, 6);
529 	immdp->op = FW_RI_DATA_IMMD;
530 	immdp->immdlen = 16;
531 	return;
532 }
533 
build_rdma_write_cmpl(struct t4_sq * sq,struct fw_ri_rdma_write_cmpl_wr * wcwr,const struct ib_send_wr * wr,u8 * len16)534 static void build_rdma_write_cmpl(struct t4_sq *sq,
535 				  struct fw_ri_rdma_write_cmpl_wr *wcwr,
536 				  const struct ib_send_wr *wr, u8 *len16)
537 {
538 	u32 plen;
539 	int size;
540 
541 	/*
542 	 * This code assumes the struct fields preceeding the write isgl
543 	 * fit in one 64B WR slot.  This is because the WQE is built
544 	 * directly in the dma queue, and wrapping is only handled
545 	 * by the code buildling sgls.  IE the "fixed part" of the wr
546 	 * structs must all fit in 64B.  The WQE build code should probably be
547 	 * redesigned to avoid this restriction, but for now just add
548 	 * the BUILD_BUG_ON() to catch if this WQE struct gets too big.
549 	 */
550 	BUILD_BUG_ON(offsetof(struct fw_ri_rdma_write_cmpl_wr, u) > 64);
551 
552 	wcwr->stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
553 	wcwr->to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
554 	if (wr->next->opcode == IB_WR_SEND)
555 		wcwr->stag_inv = 0;
556 	else
557 		wcwr->stag_inv = cpu_to_be32(wr->next->ex.invalidate_rkey);
558 	wcwr->r2 = 0;
559 	wcwr->r3 = 0;
560 
561  	/* SEND_INV SGL */
562 	if (wr->next->send_flags & IB_SEND_INLINE)
563 		build_immd_cmpl(sq, &wcwr->u_cmpl.immd_src, wr->next);
564 	else
565 		build_isgl((__be64 *)sq->queue, (__be64 *)&sq->queue[sq->size],
566 			   &wcwr->u_cmpl.isgl_src, wr->next->sg_list, 1, NULL);
567 
568 	/* WRITE SGL */
569 	build_isgl((__be64 *)sq->queue, (__be64 *)&sq->queue[sq->size],
570 		   wcwr->u.isgl_src, wr->sg_list, wr->num_sge, &plen);
571 
572 	size = sizeof *wcwr + sizeof(struct fw_ri_isgl) +
573 	       wr->num_sge * sizeof(struct fw_ri_sge);
574 	wcwr->plen = cpu_to_be32(plen);
575 	*len16 = DIV_ROUND_UP(size, 16);
576 
577 	return;
578 }
579 
build_rdma_read(union t4_wr * wqe,const struct ib_send_wr * wr,u8 * len16)580 static int build_rdma_read(union t4_wr *wqe, const struct ib_send_wr *wr, u8 *len16)
581 {
582 	if (wr->num_sge > 1)
583 		return -EINVAL;
584 	if (wr->num_sge && wr->sg_list[0].length) {
585 		wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
586 		wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
587 							>> 32));
588 		wqe->read.to_src_lo =
589 			cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
590 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
591 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
592 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
593 							 >> 32));
594 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
595 	} else {
596 		wqe->read.stag_src = cpu_to_be32(2);
597 		wqe->read.to_src_hi = 0;
598 		wqe->read.to_src_lo = 0;
599 		wqe->read.stag_sink = cpu_to_be32(2);
600 		wqe->read.plen = 0;
601 		wqe->read.to_sink_hi = 0;
602 		wqe->read.to_sink_lo = 0;
603 	}
604 	wqe->read.r2 = 0;
605 	wqe->read.r5 = 0;
606 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
607 	return 0;
608 }
609 
build_rdma_recv(struct c4iw_qp * qhp,union t4_recv_wr * wqe,const struct ib_recv_wr * wr,u8 * len16)610 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
611 			   const struct ib_recv_wr *wr, u8 *len16)
612 {
613 	int ret;
614 
615 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
616 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
617 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
618 	if (ret)
619 		return ret;
620 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
621 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
622 	return 0;
623 }
624 
build_inv_stag(union t4_wr * wqe,const struct ib_send_wr * wr,u8 * len16)625 static int build_inv_stag(union t4_wr *wqe, const struct ib_send_wr *wr,
626 			  u8 *len16)
627 {
628 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
629 	wqe->inv.r2 = 0;
630 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
631 	return 0;
632 }
633 
c4iw_qp_add_ref(struct ib_qp * qp)634 void c4iw_qp_add_ref(struct ib_qp *qp)
635 {
636 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
637 	refcount_inc(&to_c4iw_qp(qp)->qp_refcnt);
638 }
639 
c4iw_qp_rem_ref(struct ib_qp * qp)640 void c4iw_qp_rem_ref(struct ib_qp *qp)
641 {
642 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
643 	if (refcount_dec_and_test(&to_c4iw_qp(qp)->qp_refcnt))
644 		complete(&to_c4iw_qp(qp)->qp_rel_comp);
645 }
646 
ib_to_fw_opcode(int ib_opcode)647 static int ib_to_fw_opcode(int ib_opcode)
648 {
649 	int opcode;
650 
651 	switch (ib_opcode) {
652 	case IB_WR_SEND_WITH_INV:
653 		opcode = FW_RI_SEND_WITH_INV;
654 		break;
655 	case IB_WR_SEND:
656 		opcode = FW_RI_SEND;
657 		break;
658 	case IB_WR_RDMA_WRITE:
659 		opcode = FW_RI_RDMA_WRITE;
660 		break;
661 	case IB_WR_RDMA_WRITE_WITH_IMM:
662 		opcode = FW_RI_WRITE_IMMEDIATE;
663 		break;
664 	case IB_WR_RDMA_READ:
665 	case IB_WR_RDMA_READ_WITH_INV:
666 		opcode = FW_RI_READ_REQ;
667 		break;
668 	case IB_WR_REG_MR:
669 		opcode = FW_RI_FAST_REGISTER;
670 		break;
671 	case IB_WR_LOCAL_INV:
672 		opcode = FW_RI_LOCAL_INV;
673 		break;
674 	default:
675 		opcode = -EINVAL;
676 	}
677 	return opcode;
678 }
679 
complete_sq_drain_wr(struct c4iw_qp * qhp,const struct ib_send_wr * wr)680 static int complete_sq_drain_wr(struct c4iw_qp *qhp,
681 				const struct ib_send_wr *wr)
682 {
683 	struct t4_cqe cqe = {};
684 	struct c4iw_cq *schp;
685 	unsigned long flag;
686 	struct t4_cq *cq;
687 	int opcode;
688 
689 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
690 	cq = &schp->cq;
691 
692 	opcode = ib_to_fw_opcode(wr->opcode);
693 	if (opcode < 0)
694 		return opcode;
695 
696 	PDBG("%s drain sq id %u\n", __func__, qhp->wq.sq.qid);
697 	cqe.u.drain_cookie = wr->wr_id;
698 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
699 				 V_CQE_OPCODE(opcode) |
700 				 V_CQE_TYPE(1) |
701 				 V_CQE_SWCQE(1) |
702 				 V_CQE_DRAIN(1) |
703 				 V_CQE_QPID(qhp->wq.sq.qid));
704 
705 	spin_lock_irqsave(&schp->lock, flag);
706 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
707 	cq->sw_queue[cq->sw_pidx] = cqe;
708 	t4_swcq_produce(cq);
709 	spin_unlock_irqrestore(&schp->lock, flag);
710 
711 	spin_lock_irqsave(&schp->comp_handler_lock, flag);
712 	(*schp->ibcq.comp_handler)(&schp->ibcq,
713 				   schp->ibcq.cq_context);
714 	spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
715 	return 0;
716 }
717 
complete_sq_drain_wrs(struct c4iw_qp * qhp,const struct ib_send_wr * wr,const struct ib_send_wr ** bad_wr)718 static int complete_sq_drain_wrs(struct c4iw_qp *qhp, const struct ib_send_wr *wr,
719 				 const struct ib_send_wr **bad_wr)
720 {
721 	int ret = 0;
722 
723 	while (wr) {
724 		ret = complete_sq_drain_wr(qhp, wr);
725 		if (ret) {
726 			*bad_wr = wr;
727 			break;
728 		}
729 		wr = wr->next;
730 	}
731 	return ret;
732 }
733 
complete_rq_drain_wr(struct c4iw_qp * qhp,const struct ib_recv_wr * wr)734 static void complete_rq_drain_wr(struct c4iw_qp *qhp,
735 				 const struct ib_recv_wr *wr)
736 {
737 	struct t4_cqe cqe = {};
738 	struct c4iw_cq *rchp;
739 	unsigned long flag;
740 	struct t4_cq *cq;
741 
742 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
743 	cq = &rchp->cq;
744 
745 	PDBG("%s drain rq id %u\n", __func__, qhp->wq.sq.qid);
746 	cqe.u.drain_cookie = wr->wr_id;
747 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
748 				 V_CQE_OPCODE(FW_RI_SEND) |
749 				 V_CQE_TYPE(0) |
750 				 V_CQE_SWCQE(1) |
751 				 V_CQE_DRAIN(1) |
752 				 V_CQE_QPID(qhp->wq.sq.qid));
753 
754 	spin_lock_irqsave(&rchp->lock, flag);
755 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
756 	cq->sw_queue[cq->sw_pidx] = cqe;
757 	t4_swcq_produce(cq);
758 	spin_unlock_irqrestore(&rchp->lock, flag);
759 
760 	spin_lock_irqsave(&rchp->comp_handler_lock, flag);
761 	(*rchp->ibcq.comp_handler)(&rchp->ibcq,
762 				   rchp->ibcq.cq_context);
763 	spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
764 }
765 
complete_rq_drain_wrs(struct c4iw_qp * qhp,const struct ib_recv_wr * wr)766 static void complete_rq_drain_wrs(struct c4iw_qp *qhp,
767 				  const struct ib_recv_wr *wr)
768 {
769 	while (wr) {
770 		complete_rq_drain_wr(qhp, wr);
771 		wr = wr->next;
772 	}
773 }
774 
post_write_cmpl(struct c4iw_qp * qhp,const struct ib_send_wr * wr)775 static void post_write_cmpl(struct c4iw_qp *qhp, const struct ib_send_wr *wr)
776 {
777 	bool send_signaled = (wr->next->send_flags & IB_SEND_SIGNALED) ||
778 			     qhp->sq_sig_all;
779 	bool write_signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
780 			      qhp->sq_sig_all;
781 	struct t4_swsqe *swsqe;
782 	union t4_wr *wqe;
783 	u16 write_wrid;
784 	u8 len16;
785 	u16 idx;
786 
787 	/*
788 	 * The sw_sq entries still look like a WRITE and a SEND and consume
789 	 * 2 slots. The FW WR, however, will be a single uber-WR.
790 	 */
791 	wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
792 	      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
793 	build_rdma_write_cmpl(&qhp->wq.sq, &wqe->write_cmpl, wr, &len16);
794 
795 	/* WRITE swsqe */
796 	swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
797 	swsqe->opcode = FW_RI_RDMA_WRITE;
798 	swsqe->idx = qhp->wq.sq.pidx;
799 	swsqe->complete = 0;
800 	swsqe->signaled = write_signaled;
801 	swsqe->flushed = 0;
802 	swsqe->wr_id = wr->wr_id;
803 
804 	write_wrid = qhp->wq.sq.pidx;
805 
806 	/* just bump the sw_sq */
807 	qhp->wq.sq.in_use++;
808 	if (++qhp->wq.sq.pidx == qhp->wq.sq.size)
809 		qhp->wq.sq.pidx = 0;
810 
811 	/* SEND_WITH_INV swsqe */
812 	swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
813 	if (wr->next->opcode == IB_WR_SEND)
814 		swsqe->opcode = FW_RI_SEND;
815 	else
816 		swsqe->opcode = FW_RI_SEND_WITH_INV;
817 	swsqe->idx = qhp->wq.sq.pidx;
818 	swsqe->complete = 0;
819 	swsqe->signaled = send_signaled;
820 	swsqe->flushed = 0;
821 	swsqe->wr_id = wr->next->wr_id;
822 
823 	wqe->write_cmpl.flags_send = send_signaled ? FW_RI_COMPLETION_FLAG : 0;
824 	wqe->write_cmpl.wrid_send = qhp->wq.sq.pidx;
825 
826 	init_wr_hdr(wqe, write_wrid, FW_RI_RDMA_WRITE_CMPL_WR,
827 		    write_signaled ? FW_RI_COMPLETION_FLAG : 0, len16);
828 	t4_sq_produce(&qhp->wq, len16);
829 	idx = DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
830 
831 	t4_ring_sq_db(&qhp->wq, idx, wqe, qhp->rhp->rdev.adap->iwt.wc_en);
832 	return;
833 }
834 
build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr * fr,const struct ib_reg_wr * wr,struct c4iw_mr * mhp,u8 * len16)835 static int build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr,
836 		const struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16)
837 {
838 	__be64 *p = (__be64 *)fr->pbl;
839 
840 	if (wr->mr->page_size > C4IW_MAX_PAGE_SIZE)
841 		return -EINVAL;
842 
843 	fr->r2 = cpu_to_be32(0);
844 	fr->stag = cpu_to_be32(mhp->ibmr.rkey);
845 
846 	fr->tpte.valid_to_pdid = cpu_to_be32(F_FW_RI_TPTE_VALID |
847 			V_FW_RI_TPTE_STAGKEY((mhp->ibmr.rkey & M_FW_RI_TPTE_STAGKEY)) |
848 			V_FW_RI_TPTE_STAGSTATE(1) |
849 			V_FW_RI_TPTE_STAGTYPE(FW_RI_STAG_NSMR) |
850 			V_FW_RI_TPTE_PDID(mhp->attr.pdid));
851 	fr->tpte.locread_to_qpid = cpu_to_be32(
852 			V_FW_RI_TPTE_PERM(c4iw_ib_to_tpt_access(wr->access)) |
853 			V_FW_RI_TPTE_ADDRTYPE(FW_RI_VA_BASED_TO) |
854 			V_FW_RI_TPTE_PS(ilog2(wr->mr->page_size) - 12));
855 	fr->tpte.nosnoop_pbladdr = cpu_to_be32(V_FW_RI_TPTE_PBLADDR(
856 			      PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3));
857 	fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0);
858 	fr->tpte.len_hi = cpu_to_be32(mhp->ibmr.length >> 32);
859 	fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length & 0xffffffff);
860 	fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
861 	fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
862 
863 	p[0] = cpu_to_be64((u64)mhp->mpl[0]);
864 	p[1] = cpu_to_be64((u64)mhp->mpl[1]);
865 
866 	*len16 = DIV_ROUND_UP(sizeof(*fr), 16);
867 	return 0;
868 }
869 
build_memreg(struct t4_sq * sq,union t4_wr * wqe,const struct ib_reg_wr * wr,struct c4iw_mr * mhp,u8 * len16,bool dsgl_supported)870 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
871 		const struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16,
872 		bool dsgl_supported)
873 {
874 	struct fw_ri_immd *imdp;
875 	__be64 *p;
876 	int i;
877 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
878 	int rem;
879 
880 	if (mhp->mpl_len > t4_max_fr_depth(&mhp->rhp->rdev, use_dsgl))
881 		return -EINVAL;
882 	if (wr->mr->page_size > C4IW_MAX_PAGE_SIZE)
883 		return -EINVAL;
884 
885 	wqe->fr.qpbinde_to_dcacpu = 0;
886 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
887 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
888 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
889 	wqe->fr.len_hi = cpu_to_be32(mhp->ibmr.length >> 32);
890 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length & 0xffffffff);
891 	wqe->fr.stag = cpu_to_be32(wr->key);
892 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
893 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
894 
895 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
896 		struct fw_ri_dsgl *sglp;
897 
898 		for (i = 0; i < mhp->mpl_len; i++)
899 			mhp->mpl[i] =
900 				     (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
901 
902 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
903 		sglp->op = FW_RI_DATA_DSGL;
904 		sglp->r1 = 0;
905 		sglp->nsge = cpu_to_be16(1);
906 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
907 		sglp->len0 = cpu_to_be32(pbllen);
908 
909 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
910 	} else {
911 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
912 		imdp->op = FW_RI_DATA_IMMD;
913 		imdp->r1 = 0;
914 		imdp->r2 = 0;
915 		imdp->immdlen = cpu_to_be32(pbllen);
916 		p = (__be64 *)(imdp + 1);
917 		rem = pbllen;
918 		for (i = 0; i < mhp->mpl_len; i++) {
919 			*p = cpu_to_be64((u64)mhp->mpl[i]);
920 			rem -= sizeof(*p);
921 			if (++p == (__be64 *)&sq->queue[sq->size])
922 				p = (__be64 *)sq->queue;
923 		}
924 		BUG_ON(rem < 0);
925 		while (rem) {
926 			*p = 0;
927 			rem -= sizeof(*p);
928 			if (++p == (__be64 *)&sq->queue[sq->size])
929 				p = (__be64 *)sq->queue;
930 		}
931 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
932 				+ pbllen, 16);
933 	}
934 
935 	return 0;
936 }
937 
c4iw_post_send(struct ib_qp * ibqp,const struct ib_send_wr * wr,const struct ib_send_wr ** bad_wr)938 int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
939 		   const struct ib_send_wr **bad_wr)
940 {
941 	int err = 0;
942 	u8 len16 = 0;
943 	enum fw_wr_opcodes fw_opcode = 0;
944 	enum fw_ri_wr_flags fw_flags;
945 	struct c4iw_qp *qhp;
946 	union t4_wr *wqe = NULL;
947 	u32 num_wrs;
948 	struct t4_swsqe *swsqe;
949 	unsigned long flag;
950 	u16 idx = 0;
951 	struct c4iw_rdev *rdev;
952 
953 	qhp = to_c4iw_qp(ibqp);
954 	rdev = &qhp->rhp->rdev;
955 	if (__predict_false(c4iw_stopped(rdev)))
956 		return -EIO;
957 	spin_lock_irqsave(&qhp->lock, flag);
958 
959 	/*
960 	 * If the qp has been flushed, then just insert a special
961 	 * drain cqe.
962 	 */
963 	if (qhp->wq.flushed) {
964 		spin_unlock_irqrestore(&qhp->lock, flag);
965 		err = complete_sq_drain_wrs(qhp, wr, bad_wr);
966 		return err;
967 	}
968 	num_wrs = t4_sq_avail(&qhp->wq);
969 	if (num_wrs == 0) {
970 		spin_unlock_irqrestore(&qhp->lock, flag);
971 		*bad_wr = wr;
972 		return -ENOMEM;
973 	}
974 
975 	/*
976 	 * Fastpath for NVMe-oF target WRITE + SEND_WITH_INV wr chain which is
977 	 * the response for small NVMEe-oF READ requests.  If the chain is
978 	 * exactly a WRITE->SEND_WITH_INV or a WRITE->SEND and the sgl depths
979 	 * and lengths meet the requirements of the fw_ri_write_cmpl_wr work
980 	 * request, then build and post the write_cmpl WR. If any of the tests
981 	 * below are not true, then we continue on with the tradtional WRITE
982 	 * and SEND WRs.
983 	 */
984 	if (qhp->rhp->rdev.adap->params.write_cmpl_support &&
985 	    chip_id(qhp->rhp->rdev.adap) >=
986 	    CHELSIO_T5 &&
987 	    wr && wr->next && !wr->next->next &&
988 	    wr->opcode == IB_WR_RDMA_WRITE &&
989 	    wr->sg_list[0].length && wr->num_sge <= T4_WRITE_CMPL_MAX_SGL &&
990 	    (wr->next->opcode == IB_WR_SEND ||
991 	    wr->next->opcode == IB_WR_SEND_WITH_INV) &&
992 	    wr->next->num_sge == 1 && num_wrs >= 2) {
993 		post_write_cmpl(qhp, wr);
994 		spin_unlock_irqrestore(&qhp->lock, flag);
995 		return 0;
996 	}
997 
998 	while (wr) {
999 		if (num_wrs == 0) {
1000 			err = -ENOMEM;
1001 			*bad_wr = wr;
1002 			break;
1003 		}
1004 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
1005 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
1006 
1007 		fw_flags = 0;
1008 		if (wr->send_flags & IB_SEND_SOLICITED)
1009 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
1010 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
1011 			fw_flags |= FW_RI_COMPLETION_FLAG;
1012 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
1013 		switch (wr->opcode) {
1014 		case IB_WR_SEND_WITH_INV:
1015 		case IB_WR_SEND:
1016 			if (wr->send_flags & IB_SEND_FENCE)
1017 				fw_flags |= FW_RI_READ_FENCE_FLAG;
1018 			fw_opcode = FW_RI_SEND_WR;
1019 			if (wr->opcode == IB_WR_SEND)
1020 				swsqe->opcode = FW_RI_SEND;
1021 			else
1022 				swsqe->opcode = FW_RI_SEND_WITH_INV;
1023 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
1024 			break;
1025 		case IB_WR_RDMA_WRITE_WITH_IMM:
1026 			if (unlikely(!qhp->rhp->rdev.adap->params.write_w_imm_support)) {
1027 				err = -ENOSYS;
1028 				break;
1029 			}
1030 			fw_flags |= FW_RI_RDMA_WRITE_WITH_IMMEDIATE;
1031 			/*FALLTHROUGH*/
1032 		case IB_WR_RDMA_WRITE:
1033 			fw_opcode = FW_RI_RDMA_WRITE_WR;
1034 			swsqe->opcode = FW_RI_RDMA_WRITE;
1035 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
1036 			break;
1037 		case IB_WR_RDMA_READ:
1038 		case IB_WR_RDMA_READ_WITH_INV:
1039 			fw_opcode = FW_RI_RDMA_READ_WR;
1040 			swsqe->opcode = FW_RI_READ_REQ;
1041 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
1042 				c4iw_invalidate_mr(qhp->rhp,
1043 						   wr->sg_list[0].lkey);
1044 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
1045 			} else {
1046 				fw_flags = 0;
1047 			}
1048 			err = build_rdma_read(wqe, wr, &len16);
1049 			if (err)
1050 				break;
1051 			swsqe->read_len = wr->sg_list[0].length;
1052 			if (!qhp->wq.sq.oldest_read)
1053 				qhp->wq.sq.oldest_read = swsqe;
1054 			break;
1055 		case IB_WR_REG_MR: {
1056 			struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
1057 
1058 			swsqe->opcode = FW_RI_FAST_REGISTER;
1059 			if (rdev->adap->params.fr_nsmr_tpte_wr_support &&
1060 					!mhp->attr.state && mhp->mpl_len <= 2) {
1061 				fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
1062 				err = build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
1063 						mhp, &len16);
1064 			} else {
1065 				fw_opcode = FW_RI_FR_NSMR_WR;
1066 				err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
1067 					mhp, &len16,
1068 					rdev->adap->params.ulptx_memwrite_dsgl);
1069 			}
1070 			if (err)
1071 				break;
1072 			mhp->attr.state = 1;
1073 			break;
1074 		}
1075 		case IB_WR_LOCAL_INV:
1076 			if (wr->send_flags & IB_SEND_FENCE)
1077 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
1078 			fw_opcode = FW_RI_INV_LSTAG_WR;
1079 			swsqe->opcode = FW_RI_LOCAL_INV;
1080 			err = build_inv_stag(wqe, wr, &len16);
1081 			c4iw_invalidate_mr(qhp->rhp, wr->ex.invalidate_rkey);
1082 			break;
1083 		default:
1084 			CTR2(KTR_IW_CXGBE, "%s post of type =%d TBD!", __func__,
1085 			     wr->opcode);
1086 			err = -EINVAL;
1087 		}
1088 		if (err) {
1089 			*bad_wr = wr;
1090 			break;
1091 		}
1092 		swsqe->idx = qhp->wq.sq.pidx;
1093 		swsqe->complete = 0;
1094 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
1095 					qhp->sq_sig_all;
1096 		swsqe->flushed = 0;
1097 		swsqe->wr_id = wr->wr_id;
1098 
1099 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
1100 
1101 		CTR5(KTR_IW_CXGBE,
1102 		    "%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u",
1103 		    __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
1104 		    swsqe->opcode, swsqe->read_len);
1105 		wr = wr->next;
1106 		num_wrs--;
1107 		t4_sq_produce(&qhp->wq, len16);
1108 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1109 	}
1110 
1111 	t4_ring_sq_db(&qhp->wq, idx, wqe, rdev->adap->iwt.wc_en);
1112 	spin_unlock_irqrestore(&qhp->lock, flag);
1113 	return err;
1114 }
1115 
c4iw_post_receive(struct ib_qp * ibqp,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)1116 int c4iw_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1117 		      const struct ib_recv_wr **bad_wr)
1118 {
1119 	int err = 0;
1120 	struct c4iw_qp *qhp;
1121 	union t4_recv_wr *wqe = NULL;
1122 	u32 num_wrs;
1123 	u8 len16 = 0;
1124 	unsigned long flag;
1125 	u16 idx = 0;
1126 
1127 	qhp = to_c4iw_qp(ibqp);
1128 	if (__predict_false(c4iw_stopped(&qhp->rhp->rdev)))
1129 		return -EIO;
1130 	spin_lock_irqsave(&qhp->lock, flag);
1131 
1132 	/*
1133 	 * If the qp has been flushed, then just insert a special
1134 	 * drain cqe.
1135 	 */
1136 	if (qhp->wq.flushed) {
1137 		spin_unlock_irqrestore(&qhp->lock, flag);
1138 		complete_rq_drain_wrs(qhp, wr);
1139 		return err;
1140 	}
1141 	num_wrs = t4_rq_avail(&qhp->wq);
1142 	if (num_wrs == 0) {
1143 		spin_unlock_irqrestore(&qhp->lock, flag);
1144 		*bad_wr = wr;
1145 		return -ENOMEM;
1146 	}
1147 	while (wr) {
1148 		if (wr->num_sge > T4_MAX_RECV_SGE) {
1149 			err = -EINVAL;
1150 			*bad_wr = wr;
1151 			break;
1152 		}
1153 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
1154 					   qhp->wq.rq.wq_pidx *
1155 					   T4_EQ_ENTRY_SIZE);
1156 		if (num_wrs)
1157 			err = build_rdma_recv(qhp, wqe, wr, &len16);
1158 		else
1159 			err = -ENOMEM;
1160 		if (err) {
1161 			*bad_wr = wr;
1162 			break;
1163 		}
1164 
1165 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
1166 
1167 		wqe->recv.opcode = FW_RI_RECV_WR;
1168 		wqe->recv.r1 = 0;
1169 		wqe->recv.wrid = qhp->wq.rq.pidx;
1170 		wqe->recv.r2[0] = 0;
1171 		wqe->recv.r2[1] = 0;
1172 		wqe->recv.r2[2] = 0;
1173 		wqe->recv.len16 = len16;
1174 		CTR3(KTR_IW_CXGBE, "%s cookie 0x%llx pidx %u", __func__,
1175 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
1176 		t4_rq_produce(&qhp->wq, len16);
1177 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
1178 		wr = wr->next;
1179 		num_wrs--;
1180 	}
1181 
1182 	t4_ring_rq_db(&qhp->wq, idx, wqe, qhp->rhp->rdev.adap->iwt.wc_en);
1183 	spin_unlock_irqrestore(&qhp->lock, flag);
1184 	return err;
1185 }
1186 
build_term_codes(struct t4_cqe * err_cqe,u8 * layer_type,u8 * ecode)1187 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
1188 				    u8 *ecode)
1189 {
1190 	int status;
1191 	int tagged;
1192 	int opcode;
1193 	int rqtype;
1194 	int send_inv;
1195 
1196 	if (!err_cqe) {
1197 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1198 		*ecode = 0;
1199 		return;
1200 	}
1201 
1202 	status = CQE_STATUS(err_cqe);
1203 	opcode = CQE_OPCODE(err_cqe);
1204 	rqtype = RQ_TYPE(err_cqe);
1205 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
1206 		   (opcode == FW_RI_SEND_WITH_SE_INV);
1207 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
1208 		 (rqtype && (opcode == FW_RI_READ_RESP));
1209 
1210 	switch (status) {
1211 	case T4_ERR_STAG:
1212 		if (send_inv) {
1213 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1214 			*ecode = RDMAP_CANT_INV_STAG;
1215 		} else {
1216 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1217 			*ecode = RDMAP_INV_STAG;
1218 		}
1219 		break;
1220 	case T4_ERR_PDID:
1221 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1222 		if ((opcode == FW_RI_SEND_WITH_INV) ||
1223 		    (opcode == FW_RI_SEND_WITH_SE_INV))
1224 			*ecode = RDMAP_CANT_INV_STAG;
1225 		else
1226 			*ecode = RDMAP_STAG_NOT_ASSOC;
1227 		break;
1228 	case T4_ERR_QPID:
1229 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1230 		*ecode = RDMAP_STAG_NOT_ASSOC;
1231 		break;
1232 	case T4_ERR_ACCESS:
1233 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1234 		*ecode = RDMAP_ACC_VIOL;
1235 		break;
1236 	case T4_ERR_WRAP:
1237 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1238 		*ecode = RDMAP_TO_WRAP;
1239 		break;
1240 	case T4_ERR_BOUND:
1241 		if (tagged) {
1242 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1243 			*ecode = DDPT_BASE_BOUNDS;
1244 		} else {
1245 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1246 			*ecode = RDMAP_BASE_BOUNDS;
1247 		}
1248 		break;
1249 	case T4_ERR_INVALIDATE_SHARED_MR:
1250 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1251 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1252 		*ecode = RDMAP_CANT_INV_STAG;
1253 		break;
1254 	case T4_ERR_ECC:
1255 	case T4_ERR_ECC_PSTAG:
1256 	case T4_ERR_INTERNAL_ERR:
1257 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1258 		*ecode = 0;
1259 		break;
1260 	case T4_ERR_OUT_OF_RQE:
1261 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1262 		*ecode = DDPU_INV_MSN_NOBUF;
1263 		break;
1264 	case T4_ERR_PBL_ADDR_BOUND:
1265 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1266 		*ecode = DDPT_BASE_BOUNDS;
1267 		break;
1268 	case T4_ERR_CRC:
1269 		*layer_type = LAYER_MPA|DDP_LLP;
1270 		*ecode = MPA_CRC_ERR;
1271 		break;
1272 	case T4_ERR_MARKER:
1273 		*layer_type = LAYER_MPA|DDP_LLP;
1274 		*ecode = MPA_MARKER_ERR;
1275 		break;
1276 	case T4_ERR_PDU_LEN_ERR:
1277 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1278 		*ecode = DDPU_MSG_TOOBIG;
1279 		break;
1280 	case T4_ERR_DDP_VERSION:
1281 		if (tagged) {
1282 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1283 			*ecode = DDPT_INV_VERS;
1284 		} else {
1285 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1286 			*ecode = DDPU_INV_VERS;
1287 		}
1288 		break;
1289 	case T4_ERR_RDMA_VERSION:
1290 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1291 		*ecode = RDMAP_INV_VERS;
1292 		break;
1293 	case T4_ERR_OPCODE:
1294 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1295 		*ecode = RDMAP_INV_OPCODE;
1296 		break;
1297 	case T4_ERR_DDP_QUEUE_NUM:
1298 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1299 		*ecode = DDPU_INV_QN;
1300 		break;
1301 	case T4_ERR_MSN:
1302 	case T4_ERR_MSN_GAP:
1303 	case T4_ERR_MSN_RANGE:
1304 	case T4_ERR_IRD_OVERFLOW:
1305 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1306 		*ecode = DDPU_INV_MSN_RANGE;
1307 		break;
1308 	case T4_ERR_TBIT:
1309 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1310 		*ecode = 0;
1311 		break;
1312 	case T4_ERR_MO:
1313 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1314 		*ecode = DDPU_INV_MO;
1315 		break;
1316 	default:
1317 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1318 		*ecode = 0;
1319 		break;
1320 	}
1321 }
1322 
post_terminate(struct c4iw_qp * qhp,struct t4_cqe * err_cqe,gfp_t gfp)1323 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1324 			   gfp_t gfp)
1325 {
1326 	int ret;
1327 	struct fw_ri_wr *wqe;
1328 	struct terminate_message *term;
1329 	struct wrqe *wr;
1330 	struct socket *so = qhp->ep->com.so;
1331         struct inpcb *inp = sotoinpcb(so);
1332         struct tcpcb *tp = intotcpcb(inp);
1333         struct toepcb *toep = tp->t_toe;
1334 
1335 	CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
1336 	    qhp->wq.sq.qid, qhp->ep->hwtid);
1337 
1338 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1339 	if (wr == NULL)
1340 		return;
1341         wqe = wrtod(wr);
1342 
1343 	memset(wqe, 0, sizeof *wqe);
1344 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR));
1345 	wqe->flowid_len16 = cpu_to_be32(
1346 		V_FW_WR_FLOWID(qhp->ep->hwtid) |
1347 		V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1348 
1349 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1350 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1351 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1352 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1353 		term->layer_etype = qhp->attr.layer_etype;
1354 		term->ecode = qhp->attr.ecode;
1355 	} else
1356 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1357 	ret = creds(toep, tp, sizeof(*wqe));
1358 	if (ret) {
1359 		free_wrqe(wr);
1360 		return;
1361 	}
1362 	t4_wrq_tx(qhp->rhp->rdev.adap, wr);
1363 }
1364 
1365 /* Assumes qhp lock is held. */
__flush_qp(struct c4iw_qp * qhp,struct c4iw_cq * rchp,struct c4iw_cq * schp)1366 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1367 		       struct c4iw_cq *schp)
1368 {
1369 	int count;
1370 	int rq_flushed, sq_flushed;
1371 	unsigned long flag;
1372 
1373 	CTR4(KTR_IW_CXGBE, "%s qhp %p rchp %p schp %p", __func__, qhp, rchp,
1374 	    schp);
1375 
1376 	/* locking hierarchy: cq lock first, then qp lock. */
1377 	spin_lock_irqsave(&rchp->lock, flag);
1378 	spin_lock(&qhp->lock);
1379 
1380 	if (qhp->wq.flushed) {
1381 		spin_unlock(&qhp->lock);
1382 		spin_unlock_irqrestore(&rchp->lock, flag);
1383 		return;
1384 	}
1385 	qhp->wq.flushed = 1;
1386 
1387 	c4iw_flush_hw_cq(rchp, qhp);
1388 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1389 	rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1390 	spin_unlock(&qhp->lock);
1391 	spin_unlock_irqrestore(&rchp->lock, flag);
1392 
1393 	/* locking hierarchy: cq lock first, then qp lock. */
1394 	spin_lock_irqsave(&schp->lock, flag);
1395 	spin_lock(&qhp->lock);
1396 	if (schp != rchp)
1397 		c4iw_flush_hw_cq(schp, qhp);
1398 	sq_flushed = c4iw_flush_sq(qhp);
1399 	spin_unlock(&qhp->lock);
1400 	spin_unlock_irqrestore(&schp->lock, flag);
1401 
1402 	if (schp == rchp) {
1403 		if (t4_clear_cq_armed(&rchp->cq) &&
1404 		    (rq_flushed || sq_flushed)) {
1405 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1406 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1407 						   rchp->ibcq.cq_context);
1408 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1409 		}
1410 	} else {
1411 		if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1412 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1413 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1414 						   rchp->ibcq.cq_context);
1415 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1416 		}
1417 		if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1418 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1419 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1420 						   schp->ibcq.cq_context);
1421 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1422 		}
1423 	}
1424 }
1425 
flush_qp(struct c4iw_qp * qhp)1426 static void flush_qp(struct c4iw_qp *qhp)
1427 {
1428 	struct c4iw_cq *rchp, *schp;
1429 	unsigned long flag;
1430 
1431 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1432 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1433 
1434 	t4_set_wq_in_error(&qhp->wq);
1435 	if (qhp->ibqp.uobject) {
1436 
1437 		/* qhp->wq.flush is protected by qhp->mutex */
1438 		if (qhp->wq.flushed)
1439 			return;
1440 
1441 		qhp->wq.flushed = 1;
1442 		t4_set_cq_in_error(&rchp->cq);
1443 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1444 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1445 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1446 		if (schp != rchp) {
1447 			t4_set_cq_in_error(&schp->cq);
1448 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1449 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1450 					schp->ibcq.cq_context);
1451 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1452 		}
1453 		return;
1454 	}
1455 	__flush_qp(qhp, rchp, schp);
1456 }
1457 
1458 static int
rdma_fini(struct c4iw_dev * rhp,struct c4iw_qp * qhp,struct c4iw_ep * ep)1459 rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp, struct c4iw_ep *ep)
1460 {
1461 	struct c4iw_rdev *rdev = &rhp->rdev;
1462 	struct fw_ri_wr *wqe;
1463 	int ret;
1464 	struct wrqe *wr;
1465 	struct socket *so = ep->com.so;
1466         struct tcpcb *tp = intotcpcb(sotoinpcb(so));
1467         struct toepcb *toep = tp->t_toe;
1468 
1469 	KASSERT(rhp == qhp->rhp && ep == qhp->ep, ("%s: EDOOFUS", __func__));
1470 
1471 	CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp,
1472 	    qhp->wq.sq.qid, ep, ep->hwtid);
1473 
1474 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1475 	if (wr == NULL)
1476 		return (0);
1477 	wqe = wrtod(wr);
1478 
1479 	memset(wqe, 0, sizeof *wqe);
1480 
1481 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR) | F_FW_WR_COMPL);
1482 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1483 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1484 	wqe->cookie = (uintptr_t)ep->com.wr_waitp;
1485 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1486 
1487 	c4iw_init_wr_wait(ep->com.wr_waitp);
1488 
1489 	ret = creds(toep, tp, sizeof(*wqe));
1490 	if (ret) {
1491 		free_wrqe(wr);
1492 		return ret;
1493 	}
1494 
1495 	ret = c4iw_ref_send_wait(rdev, wr, ep->com.wr_waitp, ep->hwtid,
1496 			qhp->wq.sq.qid, ep->com.so, __func__);
1497 	return ret;
1498 }
1499 
build_rtr_msg(u8 p2p_type,struct fw_ri_init * init)1500 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1501 {
1502 	CTR2(KTR_IW_CXGBE, "%s p2p_type = %d", __func__, p2p_type);
1503 	memset(&init->u, 0, sizeof init->u);
1504 	switch (p2p_type) {
1505 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1506 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1507 		init->u.write.stag_sink = cpu_to_be32(1);
1508 		init->u.write.to_sink = cpu_to_be64(1);
1509 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1510 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1511 						   sizeof(struct fw_ri_immd),
1512 						   16);
1513 		break;
1514 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1515 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1516 		init->u.read.stag_src = cpu_to_be32(1);
1517 		init->u.read.to_src_lo = cpu_to_be32(1);
1518 		init->u.read.stag_sink = cpu_to_be32(1);
1519 		init->u.read.to_sink_lo = cpu_to_be32(1);
1520 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1521 		break;
1522 	}
1523 }
1524 
1525 static int
creds(struct toepcb * toep,struct tcpcb * tp,size_t wrsize)1526 creds(struct toepcb *toep, struct tcpcb *tp, size_t wrsize)
1527 {
1528 	struct ofld_tx_sdesc *txsd;
1529 
1530 	CTR3(KTR_IW_CXGBE, "%s:creB  %p %u", __func__, toep , wrsize);
1531 	INP_WLOCK(tptoinpcb(tp));
1532 	if (tp->t_flags & TF_DISCONNECTED) {
1533 		INP_WUNLOCK(tptoinpcb(tp));
1534 		return (EINVAL);
1535 	}
1536 	txsd = &toep->txsd[toep->txsd_pidx];
1537 	KASSERT(howmany(wrsize, 16) <= MAX_OFLD_TX_SDESC_CREDITS,
1538 	    ("%s: tx_credits %zu too large", __func__, howmany(wrsize, 16)));
1539 	txsd->tx_credits = howmany(wrsize, 16);
1540 	txsd->plen = 0;
1541 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
1542 			("%s: not enough credits (%d)", __func__, toep->tx_credits));
1543 	toep->tx_credits -= txsd->tx_credits;
1544 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1545 		toep->txsd_pidx = 0;
1546 	toep->txsd_avail--;
1547 	INP_WUNLOCK(tptoinpcb(tp));
1548 	CTR5(KTR_IW_CXGBE, "%s:creE  %p %u %u %u", __func__, toep ,
1549 	    txsd->tx_credits, toep->tx_credits, toep->txsd_pidx);
1550 	return (0);
1551 }
1552 
rdma_init(struct c4iw_dev * rhp,struct c4iw_qp * qhp)1553 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1554 {
1555 	struct fw_ri_wr *wqe;
1556 	int ret;
1557 	struct wrqe *wr;
1558 	struct c4iw_ep *ep = qhp->ep;
1559 	struct c4iw_rdev *rdev = &qhp->rhp->rdev;
1560 	struct adapter *sc = rdev->adap;
1561 	struct socket *so = ep->com.so;
1562         struct tcpcb *tp = intotcpcb(sotoinpcb(so));
1563         struct toepcb *toep = tp->t_toe;
1564 
1565 	CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp,
1566 	    qhp->wq.sq.qid, ep, ep->hwtid);
1567 
1568 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1569 	if (wr == NULL)
1570 		return (0);
1571 	wqe = wrtod(wr);
1572 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1573 	if (ret) {
1574 		qhp->attr.max_ird = 0;
1575 		free_wrqe(wr);
1576 		return ret;
1577 	}
1578 
1579 	memset(wqe, 0, sizeof *wqe);
1580 
1581 	wqe->op_compl = cpu_to_be32(
1582 		V_FW_WR_OP(FW_RI_WR) |
1583 		F_FW_WR_COMPL);
1584 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1585 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1586 
1587 	wqe->cookie = (uintptr_t)ep->com.wr_waitp;
1588 
1589 	wqe->u.init.type = FW_RI_TYPE_INIT;
1590 	wqe->u.init.mpareqbit_p2ptype =
1591 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1592 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1593 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1594 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1595 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1596 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1597 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1598 	if (qhp->attr.mpa_attr.crc_enabled)
1599 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1600 
1601 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1602 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1603 			    FW_RI_QP_BIND_ENABLE;
1604 	if (!qhp->ibqp.uobject)
1605 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1606 				     FW_RI_QP_STAG0_ENABLE;
1607 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1608 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1609 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1610 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1611 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1612 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1613 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1614 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1615 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1616 	wqe->u.init.iss = cpu_to_be32(ep->snd_seq);
1617 	wqe->u.init.irs = cpu_to_be32(ep->rcv_seq);
1618 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1619 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1620 	    sc->vres.rq.start);
1621 	if (qhp->attr.mpa_attr.initiator)
1622 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1623 
1624 	c4iw_init_wr_wait(ep->com.wr_waitp);
1625 
1626 	ret = creds(toep, tp, sizeof(*wqe));
1627 	if (ret) {
1628 		free_wrqe(wr);
1629 		free_ird(rhp, qhp->attr.max_ird);
1630 		return ret;
1631 	}
1632 
1633 	ret = c4iw_ref_send_wait(rdev, wr, ep->com.wr_waitp, ep->hwtid,
1634 			qhp->wq.sq.qid, ep->com.so, __func__);
1635 
1636 	toep->params.ulp_mode = ULP_MODE_RDMA;
1637 	free_ird(rhp, qhp->attr.max_ird);
1638 
1639 	return ret;
1640 }
1641 
c4iw_modify_qp(struct c4iw_dev * rhp,struct c4iw_qp * qhp,enum c4iw_qp_attr_mask mask,struct c4iw_qp_attributes * attrs,int internal)1642 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1643 		   enum c4iw_qp_attr_mask mask,
1644 		   struct c4iw_qp_attributes *attrs,
1645 		   int internal)
1646 {
1647 	int ret = 0;
1648 	struct c4iw_qp_attributes newattr = qhp->attr;
1649 	int disconnect = 0;
1650 	int terminate = 0;
1651 	int abort = 0;
1652 	int free = 0;
1653 	struct c4iw_ep *ep = NULL;
1654 
1655 	CTR5(KTR_IW_CXGBE, "%s qhp %p sqid 0x%x rqid 0x%x ep %p", __func__, qhp,
1656 	    qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep);
1657 	CTR3(KTR_IW_CXGBE, "%s state %d -> %d", __func__, qhp->attr.state,
1658 	    (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1659 
1660 	mutex_lock(&qhp->mutex);
1661 
1662 	/* Process attr changes if in IDLE */
1663 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1664 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1665 			ret = -EIO;
1666 			goto out;
1667 		}
1668 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1669 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1670 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1671 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1672 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1673 			newattr.enable_bind = attrs->enable_bind;
1674 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1675 			if (attrs->max_ord > c4iw_max_read_depth) {
1676 				ret = -EINVAL;
1677 				goto out;
1678 			}
1679 			newattr.max_ord = attrs->max_ord;
1680 		}
1681 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1682 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1683 				ret = -EINVAL;
1684 				goto out;
1685 			}
1686 			newattr.max_ird = attrs->max_ird;
1687 		}
1688 		qhp->attr = newattr;
1689 	}
1690 
1691 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1692 		goto out;
1693 	if (qhp->attr.state == attrs->next_state)
1694 		goto out;
1695 
1696 	/* Return EINPROGRESS if QP is already in transition state.
1697 	 * Eg: CLOSING->IDLE transition or *->ERROR transition.
1698 	 * This can happen while connection is switching(due to rdma_fini)
1699 	 * from iWARP/RDDP to TOE mode and any inflight RDMA RX data will
1700 	 * reach TOE driver -> TCP stack -> iWARP driver. In this way
1701 	 * iWARP driver keep receiving inflight RDMA RX data until socket
1702 	 * is closed or aborted. And if iWARP CM is in FPDU sate, then
1703 	 * it tries to put QP in TERM state and disconnects endpoint.
1704 	 * But as QP is already in transition state, this event is ignored.
1705 	 */
1706 	if ((qhp->attr.state >= C4IW_QP_STATE_ERROR) &&
1707 		(attrs->next_state == C4IW_QP_STATE_TERMINATE)) {
1708 		ret = -EINPROGRESS;
1709 		goto out;
1710 	}
1711 
1712 	switch (qhp->attr.state) {
1713 	case C4IW_QP_STATE_IDLE:
1714 		switch (attrs->next_state) {
1715 		case C4IW_QP_STATE_RTS:
1716 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1717 				ret = -EINVAL;
1718 				goto out;
1719 			}
1720 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1721 				ret = -EINVAL;
1722 				goto out;
1723 			}
1724 			qhp->attr.mpa_attr = attrs->mpa_attr;
1725 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1726 			qhp->ep = qhp->attr.llp_stream_handle;
1727 			set_state(qhp, C4IW_QP_STATE_RTS);
1728 
1729 			/*
1730 			 * Ref the endpoint here and deref when we
1731 			 * disassociate the endpoint from the QP.  This
1732 			 * happens in CLOSING->IDLE transition or *->ERROR
1733 			 * transition.
1734 			 */
1735 			c4iw_get_ep(&qhp->ep->com);
1736 			ret = rdma_init(rhp, qhp);
1737 			if (ret)
1738 				goto err;
1739 			break;
1740 		case C4IW_QP_STATE_ERROR:
1741 			set_state(qhp, C4IW_QP_STATE_ERROR);
1742 			flush_qp(qhp);
1743 			break;
1744 		default:
1745 			ret = -EINVAL;
1746 			goto out;
1747 		}
1748 		break;
1749 	case C4IW_QP_STATE_RTS:
1750 		switch (attrs->next_state) {
1751 		case C4IW_QP_STATE_CLOSING:
1752 			BUG_ON(kref_read(&qhp->ep->com.kref) < 2);
1753 			t4_set_wq_in_error(&qhp->wq);
1754 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1755 			ep = qhp->ep;
1756 			if (!internal) {
1757 				abort = 0;
1758 				disconnect = 1;
1759 				c4iw_get_ep(&qhp->ep->com);
1760 			}
1761 			ret = rdma_fini(rhp, qhp, ep);
1762 			if (ret)
1763 				goto err;
1764 			break;
1765 		case C4IW_QP_STATE_TERMINATE:
1766 			t4_set_wq_in_error(&qhp->wq);
1767 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1768 			qhp->attr.layer_etype = attrs->layer_etype;
1769 			qhp->attr.ecode = attrs->ecode;
1770 			ep = qhp->ep;
1771 			if (!internal) {
1772 				c4iw_get_ep(&qhp->ep->com);
1773 				terminate = 1;
1774 				disconnect = 1;
1775 			} else {
1776 				terminate = qhp->attr.send_term;
1777 				ret = rdma_fini(rhp, qhp, ep);
1778 				if (ret)
1779 					goto err;
1780 			}
1781 			break;
1782 		case C4IW_QP_STATE_ERROR:
1783 			t4_set_wq_in_error(&qhp->wq);
1784 			set_state(qhp, C4IW_QP_STATE_ERROR);
1785 			if (!internal) {
1786 				abort = 1;
1787 				disconnect = 1;
1788 				ep = qhp->ep;
1789 				c4iw_get_ep(&qhp->ep->com);
1790 			}
1791 			goto err;
1792 			break;
1793 		default:
1794 			ret = -EINVAL;
1795 			goto out;
1796 		}
1797 		break;
1798 	case C4IW_QP_STATE_CLOSING:
1799 
1800 		/*
1801 		 * Allow kernel users to move to ERROR for qp draining.
1802 		 */
1803 		if (!internal && (qhp->ibqp.uobject || attrs->next_state !=
1804 				  C4IW_QP_STATE_ERROR)) {
1805 			ret = -EINVAL;
1806 			goto out;
1807 		}
1808 		switch (attrs->next_state) {
1809 		case C4IW_QP_STATE_IDLE:
1810 			flush_qp(qhp);
1811 			set_state(qhp, C4IW_QP_STATE_IDLE);
1812 			qhp->attr.llp_stream_handle = NULL;
1813 			c4iw_put_ep(&qhp->ep->com);
1814 			qhp->ep = NULL;
1815 			wake_up(&qhp->wait);
1816 			break;
1817 		case C4IW_QP_STATE_ERROR:
1818 			goto err;
1819 		default:
1820 			ret = -EINVAL;
1821 			goto err;
1822 		}
1823 		break;
1824 	case C4IW_QP_STATE_ERROR:
1825 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1826 			ret = -EINVAL;
1827 			goto out;
1828 		}
1829 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1830 			ret = -EINVAL;
1831 			goto out;
1832 		}
1833 		set_state(qhp, C4IW_QP_STATE_IDLE);
1834 		break;
1835 	case C4IW_QP_STATE_TERMINATE:
1836 		if (!internal) {
1837 			ret = -EINVAL;
1838 			goto out;
1839 		}
1840 		goto err;
1841 		break;
1842 	default:
1843 		printf("%s in a bad state %d\n",
1844 		       __func__, qhp->attr.state);
1845 		ret = -EINVAL;
1846 		goto err;
1847 		break;
1848 	}
1849 	goto out;
1850 err:
1851 	CTR3(KTR_IW_CXGBE, "%s disassociating ep %p qpid 0x%x", __func__,
1852 	    qhp->ep, qhp->wq.sq.qid);
1853 
1854 	/* disassociate the LLP connection */
1855 	qhp->attr.llp_stream_handle = NULL;
1856 	if (!ep)
1857 		ep = qhp->ep;
1858 	qhp->ep = NULL;
1859 	set_state(qhp, C4IW_QP_STATE_ERROR);
1860 	free = 1;
1861 	abort = 1;
1862 	BUG_ON(!ep);
1863 	flush_qp(qhp);
1864 	wake_up(&qhp->wait);
1865 out:
1866 	mutex_unlock(&qhp->mutex);
1867 
1868 	if (terminate)
1869 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1870 
1871 	/*
1872 	 * If disconnect is 1, then we need to initiate a disconnect
1873 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1874 	 * an abnormal close (RTS/CLOSING->ERROR).
1875 	 */
1876 	if (disconnect) {
1877 		__c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1878 							 GFP_KERNEL);
1879 		c4iw_put_ep(&ep->com);
1880 	}
1881 
1882 	/*
1883 	 * If free is 1, then we've disassociated the EP from the QP
1884 	 * and we need to dereference the EP.
1885 	 */
1886 	if (free)
1887 		c4iw_put_ep(&ep->com);
1888 	CTR2(KTR_IW_CXGBE, "%s exit state %d", __func__, qhp->attr.state);
1889 	return ret;
1890 }
1891 
c4iw_destroy_qp(struct ib_qp * ib_qp,struct ib_udata * udata)1892 int c4iw_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
1893 {
1894 	struct c4iw_ucontext *ucontext;
1895 	struct c4iw_dev *rhp;
1896 	struct c4iw_qp *qhp;
1897 	struct c4iw_qp_attributes attrs;
1898 
1899 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ib_qp);
1900 	qhp = to_c4iw_qp(ib_qp);
1901 	ucontext = qhp->ucontext;
1902 	rhp = qhp->rhp;
1903 
1904 	attrs.next_state = C4IW_QP_STATE_ERROR;
1905 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1906 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1907 	else
1908 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1909 	wait_event(qhp->wait, !qhp->ep);
1910 
1911 	xa_lock_irq(&rhp->qps);
1912 	__xa_erase(&rhp->qps, qhp->wq.sq.qid);
1913 	xa_unlock_irq(&rhp->qps);
1914 
1915 	free_ird(rhp, qhp->attr.max_ird);
1916 	c4iw_qp_rem_ref(ib_qp);
1917 
1918 	wait_for_completion(&qhp->qp_rel_comp);
1919 
1920 	CTR3(KTR_IW_CXGBE, "%s ib_qp %p qpid 0x%0x", __func__, ib_qp,
1921 	    qhp->wq.sq.qid);
1922 	CTR3(KTR_IW_CXGBE, "%s qhp %p ucontext %p", __func__,
1923 			qhp, ucontext);
1924 	destroy_qp(&rhp->rdev, &qhp->wq,
1925 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1926 
1927 	c4iw_put_wr_wait(qhp->wr_waitp);
1928 
1929 	kfree(qhp);
1930 
1931 	return 0;
1932 }
1933 
1934 struct ib_qp *
c4iw_create_qp(struct ib_pd * pd,struct ib_qp_init_attr * attrs,struct ib_udata * udata)1935 c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1936     struct ib_udata *udata)
1937 {
1938 	struct c4iw_dev *rhp;
1939 	struct c4iw_qp *qhp;
1940 	struct c4iw_pd *php;
1941 	struct c4iw_cq *schp;
1942 	struct c4iw_cq *rchp;
1943 	struct c4iw_create_qp_resp uresp;
1944 	unsigned int sqsize, rqsize;
1945 	struct c4iw_ucontext *ucontext = rdma_udata_to_drv_context(
1946 			udata, struct c4iw_ucontext, ibucontext);
1947 	int ret;
1948 	struct c4iw_mm_entry *sq_key_mm = NULL, *rq_key_mm = NULL;
1949 	struct c4iw_mm_entry *sq_db_key_mm = NULL, *rq_db_key_mm = NULL;
1950 
1951 	CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
1952 
1953 	if (attrs->qp_type != IB_QPT_RC)
1954 		return ERR_PTR(-EINVAL);
1955 
1956 	php = to_c4iw_pd(pd);
1957 	rhp = php->rhp;
1958 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1959 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1960 	if (!schp || !rchp)
1961 		return ERR_PTR(-EINVAL);
1962 
1963 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1964 		return ERR_PTR(-EINVAL);
1965 
1966 	if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1967 		return ERR_PTR(-E2BIG);
1968 	rqsize = attrs->cap.max_recv_wr + 1;
1969 	if (rqsize < 8)
1970 		rqsize = 8;
1971 
1972 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1973 		return ERR_PTR(-E2BIG);
1974 	sqsize = attrs->cap.max_send_wr + 1;
1975 	if (sqsize < 8)
1976 		sqsize = 8;
1977 
1978 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1979 	if (!qhp)
1980 		return ERR_PTR(-ENOMEM);
1981 	qhp->wr_waitp = c4iw_alloc_wr_wait(GFP_KERNEL);
1982 	if (!qhp->wr_waitp) {
1983 		ret = -ENOMEM;
1984 		goto err_free_qhp;
1985 	}
1986 
1987 	qhp->wq.sq.size = sqsize;
1988 	qhp->wq.sq.memsize =
1989 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1990 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1991 	qhp->wq.sq.flush_cidx = -1;
1992 	qhp->wq.rq.size = rqsize;
1993 	qhp->wq.rq.memsize =
1994 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1995 		sizeof(*qhp->wq.rq.queue);
1996 
1997 	if (ucontext) {
1998 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1999 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
2000 	}
2001 
2002 	CTR5(KTR_IW_CXGBE, "%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu",
2003 	    __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
2004 
2005 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
2006 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx,
2007 			qhp->wr_waitp);
2008 	if (ret)
2009 		goto err_free_wr_wait;
2010 
2011 	attrs->cap.max_recv_wr = rqsize - 1;
2012 	attrs->cap.max_send_wr = sqsize - 1;
2013 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
2014 
2015 	qhp->rhp = rhp;
2016 	qhp->attr.pd = php->pdid;
2017 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
2018 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
2019 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
2020 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
2021 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
2022 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
2023 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
2024 	qhp->attr.state = C4IW_QP_STATE_IDLE;
2025 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
2026 	qhp->attr.enable_rdma_read = 1;
2027 	qhp->attr.enable_rdma_write = 1;
2028 	qhp->attr.enable_bind = 1;
2029 	qhp->attr.max_ord = 0;
2030 	qhp->attr.max_ird = 0;
2031 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
2032 	spin_lock_init(&qhp->lock);
2033 	mutex_init(&qhp->mutex);
2034 	init_waitqueue_head(&qhp->wait);
2035 	init_completion(&qhp->qp_rel_comp);
2036 	refcount_set(&qhp->qp_refcnt, 1);
2037 
2038 	ret = xa_insert_irq(&rhp->qps, qhp->wq.sq.qid, qhp, GFP_KERNEL);
2039 	if (ret)
2040 		goto err_destroy_qp;
2041 
2042 	if (udata) {
2043 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
2044 		if (!sq_key_mm) {
2045 			ret = -ENOMEM;
2046 			goto err_remove_handle;
2047 		}
2048 		rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
2049 		if (!rq_key_mm) {
2050 			ret = -ENOMEM;
2051 			goto err_free_sq_key;
2052 		}
2053 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
2054 		if (!sq_db_key_mm) {
2055 			ret = -ENOMEM;
2056 			goto err_free_rq_key;
2057 		}
2058 		rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
2059 		if (!rq_db_key_mm) {
2060 			ret = -ENOMEM;
2061 			goto err_free_sq_db_key;
2062 		}
2063 		uresp.flags = 0;
2064 		if (rhp->rdev.adap->params.write_w_imm_support)
2065 			uresp.flags |= C4IW_QPF_WRITE_W_IMM;
2066 		uresp.qid_mask = rhp->rdev.qpmask;
2067 		uresp.sqid = qhp->wq.sq.qid;
2068 		uresp.sq_size = qhp->wq.sq.size;
2069 		uresp.sq_memsize = qhp->wq.sq.memsize;
2070 		uresp.rqid = qhp->wq.rq.qid;
2071 		uresp.rq_size = qhp->wq.rq.size;
2072 		uresp.rq_memsize = qhp->wq.rq.memsize;
2073 		spin_lock(&ucontext->mmap_lock);
2074 		uresp.ma_sync_key =  0;
2075 		uresp.sq_key = ucontext->key;
2076 		ucontext->key += PAGE_SIZE;
2077 		uresp.rq_key = ucontext->key;
2078 		ucontext->key += PAGE_SIZE;
2079 		uresp.sq_db_gts_key = ucontext->key;
2080 		ucontext->key += PAGE_SIZE;
2081 		uresp.rq_db_gts_key = ucontext->key;
2082 		ucontext->key += PAGE_SIZE;
2083 		spin_unlock(&ucontext->mmap_lock);
2084 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
2085 		if (ret)
2086 			goto err_free_rq_db_key;
2087 		sq_key_mm->key = uresp.sq_key;
2088 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
2089 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
2090 		CTR4(KTR_IW_CXGBE, "%s sq_key_mm %x, %x, %d", __func__,
2091 				sq_key_mm->key, sq_key_mm->addr,
2092 				sq_key_mm->len);
2093 		insert_mmap(ucontext, sq_key_mm);
2094 		rq_key_mm->key = uresp.rq_key;
2095 		rq_key_mm->addr = qhp->wq.rq.phys_addr;
2096 		rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
2097 		CTR4(KTR_IW_CXGBE, "%s rq_key_mm %x, %x, %d", __func__,
2098 				rq_key_mm->key, rq_key_mm->addr,
2099 				rq_key_mm->len);
2100 		insert_mmap(ucontext, rq_key_mm);
2101 		sq_db_key_mm->key = uresp.sq_db_gts_key;
2102 		sq_db_key_mm->addr = (u64)qhp->wq.sq.bar2_pa;
2103 		sq_db_key_mm->len = PAGE_SIZE;
2104 		CTR4(KTR_IW_CXGBE, "%s sq_db_key_mm %x, %x, %d", __func__,
2105 				sq_db_key_mm->key, sq_db_key_mm->addr,
2106 				sq_db_key_mm->len);
2107 		insert_mmap(ucontext, sq_db_key_mm);
2108 		rq_db_key_mm->key = uresp.rq_db_gts_key;
2109 		rq_db_key_mm->addr = (u64)qhp->wq.rq.bar2_pa;
2110 		rq_db_key_mm->len = PAGE_SIZE;
2111 		CTR4(KTR_IW_CXGBE, "%s rq_db_key_mm %x, %x, %d", __func__,
2112 				rq_db_key_mm->key, rq_db_key_mm->addr,
2113 				rq_db_key_mm->len);
2114 		insert_mmap(ucontext, rq_db_key_mm);
2115 
2116 		qhp->ucontext = ucontext;
2117 	}
2118 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
2119 
2120 	CTR5(KTR_IW_CXGBE, "%s sq id %u size %u memsize %zu num_entries %u",
2121 		 __func__, qhp->wq.sq.qid,
2122 		 qhp->wq.sq.size, qhp->wq.sq.memsize, attrs->cap.max_send_wr);
2123 	CTR5(KTR_IW_CXGBE, "%s rq id %u size %u memsize %zu num_entries %u",
2124 		 __func__, qhp->wq.rq.qid,
2125 		 qhp->wq.rq.size, qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
2126 	return &qhp->ibqp;
2127 err_free_rq_db_key:
2128 	kfree(rq_db_key_mm);
2129 err_free_sq_db_key:
2130 	kfree(sq_db_key_mm);
2131 err_free_rq_key:
2132 	kfree(rq_key_mm);
2133 err_free_sq_key:
2134 	kfree(sq_key_mm);
2135 err_remove_handle:
2136 	xa_erase_irq(&rhp->qps, qhp->wq.sq.qid);
2137 err_destroy_qp:
2138 	destroy_qp(&rhp->rdev, &qhp->wq,
2139 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
2140 err_free_wr_wait:
2141 	c4iw_put_wr_wait(qhp->wr_waitp);
2142 err_free_qhp:
2143 	kfree(qhp);
2144 	return ERR_PTR(ret);
2145 }
2146 
c4iw_ib_modify_qp(struct ib_qp * ibqp,struct ib_qp_attr * attr,int attr_mask,struct ib_udata * udata)2147 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2148 		      int attr_mask, struct ib_udata *udata)
2149 {
2150 	struct c4iw_dev *rhp;
2151 	struct c4iw_qp *qhp;
2152 	enum c4iw_qp_attr_mask mask = 0;
2153 	struct c4iw_qp_attributes attrs;
2154 
2155 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ibqp);
2156 
2157 	/* iwarp does not support the RTR state */
2158 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
2159 		attr_mask &= ~IB_QP_STATE;
2160 
2161 	/* Make sure we still have something left to do */
2162 	if (!attr_mask)
2163 		return 0;
2164 
2165 	memset(&attrs, 0, sizeof attrs);
2166 	qhp = to_c4iw_qp(ibqp);
2167 	rhp = qhp->rhp;
2168 
2169 	attrs.next_state = c4iw_convert_state(attr->qp_state);
2170 	attrs.enable_rdma_read = (attr->qp_access_flags &
2171 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
2172 	attrs.enable_rdma_write = (attr->qp_access_flags &
2173 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
2174 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
2175 
2176 
2177 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
2178 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
2179 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
2180 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
2181 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
2182 
2183 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
2184 }
2185 
c4iw_get_qp(struct ib_device * dev,int qpn)2186 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
2187 {
2188 	CTR3(KTR_IW_CXGBE, "%s ib_dev %p qpn 0x%x", __func__, dev, qpn);
2189 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
2190 }
2191 
c4iw_ib_query_qp(struct ib_qp * ibqp,struct ib_qp_attr * attr,int attr_mask,struct ib_qp_init_attr * init_attr)2192 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2193 		     int attr_mask, struct ib_qp_init_attr *init_attr)
2194 {
2195 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
2196 
2197 	memset(attr, 0, sizeof *attr);
2198 	memset(init_attr, 0, sizeof *init_attr);
2199 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
2200 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
2201 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
2202 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
2203 	init_attr->cap.max_recv_sge = qhp->attr.rq_max_sges;
2204 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
2205 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
2206 	return 0;
2207 }
2208 #endif
2209