xref: /freebsd/sys/dev/mthca/mthca_srq.c (revision b633e08c705fe43180567eae26923d6f6f98c8d9)
1 /*
2  * Copyright (c) 2005 Cisco Systems. All rights reserved.
3  *
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:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
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.
22  *
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
30  * SOFTWARE.
31  */
32 
33 #include <linux/slab.h>
34 #include <linux/string.h>
35 #include <linux/sched.h>
36 
37 #include <asm/io.h>
38 
39 #include <rdma/uverbs_ioctl.h>
40 
41 #include "mthca_dev.h"
42 #include "mthca_cmd.h"
43 #include "mthca_memfree.h"
44 #include "mthca_wqe.h"
45 
46 enum {
47 	MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
48 };
49 
50 struct mthca_tavor_srq_context {
51 	__be64 wqe_base_ds;	/* low 6 bits is descriptor size */
52 	__be32 state_pd;
53 	__be32 lkey;
54 	__be32 uar;
55 	__be16 limit_watermark;
56 	__be16 wqe_cnt;
57 	u32    reserved[2];
58 };
59 
60 struct mthca_arbel_srq_context {
61 	__be32 state_logsize_srqn;
62 	__be32 lkey;
63 	__be32 db_index;
64 	__be32 logstride_usrpage;
65 	__be64 wqe_base;
66 	__be32 eq_pd;
67 	__be16 limit_watermark;
68 	__be16 wqe_cnt;
69 	u16    reserved1;
70 	__be16 wqe_counter;
71 	u32    reserved2[3];
72 };
73 
get_wqe(struct mthca_srq * srq,int n)74 static void *get_wqe(struct mthca_srq *srq, int n)
75 {
76 	if (srq->is_direct)
77 		return srq->queue.direct.buf + (n << srq->wqe_shift);
78 	else
79 		return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
80 			((n << srq->wqe_shift) & (PAGE_SIZE - 1));
81 }
82 
83 /*
84  * Return a pointer to the location within a WQE that we're using as a
85  * link when the WQE is in the free list.  We use the imm field
86  * because in the Tavor case, posting a WQE may overwrite the next
87  * segment of the previous WQE, but a receive WQE will never touch the
88  * imm field.  This avoids corrupting our free list if the previous
89  * WQE has already completed and been put on the free list when we
90  * post the next WQE.
91  */
wqe_to_link(void * wqe)92 static inline int *wqe_to_link(void *wqe)
93 {
94 	return (int *) (wqe + offsetof(struct mthca_next_seg, imm));
95 }
96 
mthca_tavor_init_srq_context(struct mthca_dev * dev,struct mthca_pd * pd,struct mthca_srq * srq,struct mthca_tavor_srq_context * context,struct ib_udata * udata)97 static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
98 					 struct mthca_pd *pd,
99 					 struct mthca_srq *srq,
100 					 struct mthca_tavor_srq_context *context,
101 					 struct ib_udata *udata)
102 {
103 	struct mthca_ucontext *ucontext = rdma_udata_to_drv_context(
104 		udata, struct mthca_ucontext, ibucontext);
105 
106 	memset(context, 0, sizeof *context);
107 
108 	context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
109 	context->state_pd    = cpu_to_be32(pd->pd_num);
110 	context->lkey        = cpu_to_be32(srq->mr.ibmr.lkey);
111 
112 	if (udata)
113 		context->uar = cpu_to_be32(ucontext->uar.index);
114 	else
115 		context->uar = cpu_to_be32(dev->driver_uar.index);
116 }
117 
mthca_arbel_init_srq_context(struct mthca_dev * dev,struct mthca_pd * pd,struct mthca_srq * srq,struct mthca_arbel_srq_context * context,struct ib_udata * udata)118 static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
119 					 struct mthca_pd *pd,
120 					 struct mthca_srq *srq,
121 					 struct mthca_arbel_srq_context *context,
122 					 struct ib_udata *udata)
123 {
124 	struct mthca_ucontext *ucontext = rdma_udata_to_drv_context(
125 		udata, struct mthca_ucontext, ibucontext);
126 	int logsize;
127 
128 	memset(context, 0, sizeof *context);
129 	logsize = ilog2(srq->max);
130 	context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
131 	context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
132 	context->db_index = cpu_to_be32(srq->db_index);
133 	context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
134 	if (udata)
135 		context->logstride_usrpage |= cpu_to_be32(ucontext->uar.index);
136 	else
137 		context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
138 	context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
139 }
140 
mthca_free_srq_buf(struct mthca_dev * dev,struct mthca_srq * srq)141 static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
142 {
143 	mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
144 		       srq->is_direct, &srq->mr);
145 	kfree(srq->wrid);
146 }
147 
mthca_alloc_srq_buf(struct mthca_dev * dev,struct mthca_pd * pd,struct mthca_srq * srq,struct ib_udata * udata)148 static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
149 			       struct mthca_srq *srq, struct ib_udata *udata)
150 {
151 	struct mthca_data_seg *scatter;
152 	void *wqe;
153 	int err;
154 	int i;
155 
156 	if (udata)
157 		return 0;
158 
159 	srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
160 	if (!srq->wrid)
161 		return -ENOMEM;
162 
163 	err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
164 			      MTHCA_MAX_DIRECT_SRQ_SIZE,
165 			      &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
166 	if (err) {
167 		kfree(srq->wrid);
168 		return err;
169 	}
170 
171 	/*
172 	 * Now initialize the SRQ buffer so that all of the WQEs are
173 	 * linked into the list of free WQEs.  In addition, set the
174 	 * scatter list L_Keys to the sentry value of 0x100.
175 	 */
176 	for (i = 0; i < srq->max; ++i) {
177 		struct mthca_next_seg *next;
178 
179 		next = wqe = get_wqe(srq, i);
180 
181 		if (i < srq->max - 1) {
182 			*wqe_to_link(wqe) = i + 1;
183 			next->nda_op = htonl(((i + 1) << srq->wqe_shift) | 1);
184 		} else {
185 			*wqe_to_link(wqe) = -1;
186 			next->nda_op = 0;
187 		}
188 
189 		for (scatter = wqe + sizeof (struct mthca_next_seg);
190 		     (void *) scatter < wqe + (1 << srq->wqe_shift);
191 		     ++scatter)
192 			scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
193 	}
194 
195 	srq->last = get_wqe(srq, srq->max - 1);
196 
197 	return 0;
198 }
199 
mthca_alloc_srq(struct mthca_dev * dev,struct mthca_pd * pd,struct ib_srq_attr * attr,struct mthca_srq * srq,struct ib_udata * udata)200 int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
201 		    struct ib_srq_attr *attr, struct mthca_srq *srq,
202 		    struct ib_udata *udata)
203 {
204 	struct mthca_mailbox *mailbox;
205 	int ds;
206 	int err;
207 
208 	/* Sanity check SRQ size before proceeding */
209 	if (attr->max_wr  > dev->limits.max_srq_wqes ||
210 	    attr->max_sge > dev->limits.max_srq_sge)
211 		return -EINVAL;
212 
213 	srq->max      = attr->max_wr;
214 	srq->max_gs   = attr->max_sge;
215 	srq->counter  = 0;
216 
217 	if (mthca_is_memfree(dev))
218 		srq->max = roundup_pow_of_two(srq->max + 1);
219 	else
220 		srq->max = srq->max + 1;
221 
222 	ds = max(64UL,
223 		 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
224 				    srq->max_gs * sizeof (struct mthca_data_seg)));
225 
226 	if (!mthca_is_memfree(dev) && (ds > dev->limits.max_desc_sz))
227 		return -EINVAL;
228 
229 	srq->wqe_shift = ilog2(ds);
230 
231 	srq->srqn = mthca_alloc(&dev->srq_table.alloc);
232 	if (srq->srqn == -1)
233 		return -ENOMEM;
234 
235 	if (mthca_is_memfree(dev)) {
236 		err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
237 		if (err)
238 			goto err_out;
239 
240 		if (!udata) {
241 			srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
242 						       srq->srqn, &srq->db);
243 			if (srq->db_index < 0) {
244 				err = -ENOMEM;
245 				goto err_out_icm;
246 			}
247 		}
248 	}
249 
250 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
251 	if (IS_ERR(mailbox)) {
252 		err = PTR_ERR(mailbox);
253 		goto err_out_db;
254 	}
255 
256 	err = mthca_alloc_srq_buf(dev, pd, srq, udata);
257 	if (err)
258 		goto err_out_mailbox;
259 
260 	spin_lock_init(&srq->lock);
261 	srq->refcount = 1;
262 	init_waitqueue_head(&srq->wait);
263 	mutex_init(&srq->mutex);
264 
265 	if (mthca_is_memfree(dev))
266 		mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf, udata);
267 	else
268 		mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf, udata);
269 
270 	err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn);
271 
272 	if (err) {
273 		mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
274 		goto err_out_free_buf;
275 	}
276 
277 	spin_lock_irq(&dev->srq_table.lock);
278 	if (mthca_array_set(&dev->srq_table.srq,
279 			    srq->srqn & (dev->limits.num_srqs - 1),
280 			    srq)) {
281 		spin_unlock_irq(&dev->srq_table.lock);
282 		goto err_out_free_srq;
283 	}
284 	spin_unlock_irq(&dev->srq_table.lock);
285 
286 	mthca_free_mailbox(dev, mailbox);
287 
288 	srq->first_free = 0;
289 	srq->last_free  = srq->max - 1;
290 
291 	attr->max_wr    = srq->max - 1;
292 	attr->max_sge   = srq->max_gs;
293 
294 	return 0;
295 
296 err_out_free_srq:
297 	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
298 	if (err)
299 		mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
300 
301 err_out_free_buf:
302 	if (!udata)
303 		mthca_free_srq_buf(dev, srq);
304 
305 err_out_mailbox:
306 	mthca_free_mailbox(dev, mailbox);
307 
308 err_out_db:
309 	if (!udata && mthca_is_memfree(dev))
310 		mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
311 
312 err_out_icm:
313 	mthca_table_put(dev, dev->srq_table.table, srq->srqn);
314 
315 err_out:
316 	mthca_free(&dev->srq_table.alloc, srq->srqn);
317 
318 	return err;
319 }
320 
get_srq_refcount(struct mthca_dev * dev,struct mthca_srq * srq)321 static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
322 {
323 	int c;
324 
325 	spin_lock_irq(&dev->srq_table.lock);
326 	c = srq->refcount;
327 	spin_unlock_irq(&dev->srq_table.lock);
328 
329 	return c;
330 }
331 
mthca_free_srq(struct mthca_dev * dev,struct mthca_srq * srq)332 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
333 {
334 	struct mthca_mailbox *mailbox;
335 	int err;
336 
337 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
338 	if (IS_ERR(mailbox)) {
339 		mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
340 		return;
341 	}
342 
343 	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
344 	if (err)
345 		mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
346 
347 	spin_lock_irq(&dev->srq_table.lock);
348 	mthca_array_clear(&dev->srq_table.srq,
349 			  srq->srqn & (dev->limits.num_srqs - 1));
350 	--srq->refcount;
351 	spin_unlock_irq(&dev->srq_table.lock);
352 
353 	wait_event(srq->wait, !get_srq_refcount(dev, srq));
354 
355 	if (!srq->ibsrq.uobject) {
356 		mthca_free_srq_buf(dev, srq);
357 		if (mthca_is_memfree(dev))
358 			mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
359 	}
360 
361 	mthca_table_put(dev, dev->srq_table.table, srq->srqn);
362 	mthca_free(&dev->srq_table.alloc, srq->srqn);
363 	mthca_free_mailbox(dev, mailbox);
364 }
365 
mthca_modify_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr,enum ib_srq_attr_mask attr_mask,struct ib_udata * udata)366 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
367 		     enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
368 {
369 	struct mthca_dev *dev = to_mdev(ibsrq->device);
370 	struct mthca_srq *srq = to_msrq(ibsrq);
371 	int ret = 0;
372 
373 	/* We don't support resizing SRQs (yet?) */
374 	if (attr_mask & IB_SRQ_MAX_WR)
375 		return -EINVAL;
376 
377 	if (attr_mask & IB_SRQ_LIMIT) {
378 		u32 max_wr = mthca_is_memfree(dev) ? srq->max - 1 : srq->max;
379 		if (attr->srq_limit > max_wr)
380 			return -EINVAL;
381 
382 		mutex_lock(&srq->mutex);
383 		ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit);
384 		mutex_unlock(&srq->mutex);
385 	}
386 
387 	return ret;
388 }
389 
mthca_query_srq(struct ib_srq * ibsrq,struct ib_srq_attr * srq_attr)390 int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
391 {
392 	struct mthca_dev *dev = to_mdev(ibsrq->device);
393 	struct mthca_srq *srq = to_msrq(ibsrq);
394 	struct mthca_mailbox *mailbox;
395 	struct mthca_arbel_srq_context *arbel_ctx;
396 	struct mthca_tavor_srq_context *tavor_ctx;
397 	int err;
398 
399 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
400 	if (IS_ERR(mailbox))
401 		return PTR_ERR(mailbox);
402 
403 	err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox);
404 	if (err)
405 		goto out;
406 
407 	if (mthca_is_memfree(dev)) {
408 		arbel_ctx = mailbox->buf;
409 		srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
410 	} else {
411 		tavor_ctx = mailbox->buf;
412 		srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
413 	}
414 
415 	srq_attr->max_wr  = srq->max - 1;
416 	srq_attr->max_sge = srq->max_gs;
417 
418 out:
419 	mthca_free_mailbox(dev, mailbox);
420 
421 	return err;
422 }
423 
mthca_srq_event(struct mthca_dev * dev,u32 srqn,enum ib_event_type event_type)424 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
425 		     enum ib_event_type event_type)
426 {
427 	struct mthca_srq *srq;
428 	struct ib_event event;
429 
430 	spin_lock(&dev->srq_table.lock);
431 	srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
432 	if (srq)
433 		++srq->refcount;
434 	spin_unlock(&dev->srq_table.lock);
435 
436 	if (!srq) {
437 		mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
438 		return;
439 	}
440 
441 	if (!srq->ibsrq.event_handler)
442 		goto out;
443 
444 	event.device      = &dev->ib_dev;
445 	event.event       = event_type;
446 	event.element.srq = &srq->ibsrq;
447 	srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
448 
449 out:
450 	spin_lock(&dev->srq_table.lock);
451 	if (!--srq->refcount)
452 		wake_up(&srq->wait);
453 	spin_unlock(&dev->srq_table.lock);
454 }
455 
456 /*
457  * This function must be called with IRQs disabled.
458  */
mthca_free_srq_wqe(struct mthca_srq * srq,u32 wqe_addr)459 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
460 {
461 	int ind;
462 	struct mthca_next_seg *last_free;
463 
464 	ind = wqe_addr >> srq->wqe_shift;
465 
466 	spin_lock(&srq->lock);
467 
468 	last_free = get_wqe(srq, srq->last_free);
469 	*wqe_to_link(last_free) = ind;
470 	last_free->nda_op = htonl((ind << srq->wqe_shift) | 1);
471 	*wqe_to_link(get_wqe(srq, ind)) = -1;
472 	srq->last_free = ind;
473 
474 	spin_unlock(&srq->lock);
475 }
476 
mthca_tavor_post_srq_recv(struct ib_srq * ibsrq,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)477 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
478 			      const struct ib_recv_wr **bad_wr)
479 {
480 	struct mthca_dev *dev = to_mdev(ibsrq->device);
481 	struct mthca_srq *srq = to_msrq(ibsrq);
482 	unsigned long flags;
483 	int err = 0;
484 	int first_ind;
485 	int ind;
486 	int next_ind;
487 	int nreq;
488 	int i;
489 	void *wqe;
490 	void *prev_wqe;
491 
492 	spin_lock_irqsave(&srq->lock, flags);
493 
494 	first_ind = srq->first_free;
495 
496 	for (nreq = 0; wr; wr = wr->next) {
497 		ind       = srq->first_free;
498 		wqe       = get_wqe(srq, ind);
499 		next_ind  = *wqe_to_link(wqe);
500 
501 		if (unlikely(next_ind < 0)) {
502 			mthca_err(dev, "SRQ %06x full\n", srq->srqn);
503 			err = -ENOMEM;
504 			*bad_wr = wr;
505 			break;
506 		}
507 
508 		prev_wqe  = srq->last;
509 		srq->last = wqe;
510 
511 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
512 		/* flags field will always remain 0 */
513 
514 		wqe += sizeof (struct mthca_next_seg);
515 
516 		if (unlikely(wr->num_sge > srq->max_gs)) {
517 			err = -EINVAL;
518 			*bad_wr = wr;
519 			srq->last = prev_wqe;
520 			break;
521 		}
522 
523 		for (i = 0; i < wr->num_sge; ++i) {
524 			mthca_set_data_seg(wqe, wr->sg_list + i);
525 			wqe += sizeof (struct mthca_data_seg);
526 		}
527 
528 		if (i < srq->max_gs)
529 			mthca_set_data_seg_inval(wqe);
530 
531 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
532 			cpu_to_be32(MTHCA_NEXT_DBD);
533 
534 		srq->wrid[ind]  = wr->wr_id;
535 		srq->first_free = next_ind;
536 
537 		++nreq;
538 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
539 			nreq = 0;
540 
541 			/*
542 			 * Make sure that descriptors are written
543 			 * before doorbell is rung.
544 			 */
545 			wmb();
546 
547 			mthca_write64(first_ind << srq->wqe_shift, srq->srqn << 8,
548 				      dev->kar + MTHCA_RECEIVE_DOORBELL,
549 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
550 
551 			first_ind = srq->first_free;
552 		}
553 	}
554 
555 	if (likely(nreq)) {
556 		/*
557 		 * Make sure that descriptors are written before
558 		 * doorbell is rung.
559 		 */
560 		wmb();
561 
562 		mthca_write64(first_ind << srq->wqe_shift, (srq->srqn << 8) | nreq,
563 			      dev->kar + MTHCA_RECEIVE_DOORBELL,
564 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
565 	}
566 
567 	/*
568 	 * Make sure doorbells don't leak out of SRQ spinlock and
569 	 * reach the HCA out of order:
570 	 */
571 	mmiowb();
572 
573 	spin_unlock_irqrestore(&srq->lock, flags);
574 	return err;
575 }
576 
mthca_arbel_post_srq_recv(struct ib_srq * ibsrq,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)577 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
578 			      const struct ib_recv_wr **bad_wr)
579 {
580 	struct mthca_dev *dev = to_mdev(ibsrq->device);
581 	struct mthca_srq *srq = to_msrq(ibsrq);
582 	unsigned long flags;
583 	int err = 0;
584 	int ind;
585 	int next_ind;
586 	int nreq;
587 	int i;
588 	void *wqe;
589 
590 	spin_lock_irqsave(&srq->lock, flags);
591 
592 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
593 		ind       = srq->first_free;
594 		wqe       = get_wqe(srq, ind);
595 		next_ind  = *wqe_to_link(wqe);
596 
597 		if (unlikely(next_ind < 0)) {
598 			mthca_err(dev, "SRQ %06x full\n", srq->srqn);
599 			err = -ENOMEM;
600 			*bad_wr = wr;
601 			break;
602 		}
603 
604 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
605 		/* flags field will always remain 0 */
606 
607 		wqe += sizeof (struct mthca_next_seg);
608 
609 		if (unlikely(wr->num_sge > srq->max_gs)) {
610 			err = -EINVAL;
611 			*bad_wr = wr;
612 			break;
613 		}
614 
615 		for (i = 0; i < wr->num_sge; ++i) {
616 			mthca_set_data_seg(wqe, wr->sg_list + i);
617 			wqe += sizeof (struct mthca_data_seg);
618 		}
619 
620 		if (i < srq->max_gs)
621 			mthca_set_data_seg_inval(wqe);
622 
623 		srq->wrid[ind]  = wr->wr_id;
624 		srq->first_free = next_ind;
625 	}
626 
627 	if (likely(nreq)) {
628 		srq->counter += nreq;
629 
630 		/*
631 		 * Make sure that descriptors are written before
632 		 * we write doorbell record.
633 		 */
634 		wmb();
635 		*srq->db = cpu_to_be32(srq->counter);
636 	}
637 
638 	spin_unlock_irqrestore(&srq->lock, flags);
639 	return err;
640 }
641 
mthca_max_srq_sge(struct mthca_dev * dev)642 int mthca_max_srq_sge(struct mthca_dev *dev)
643 {
644 	if (mthca_is_memfree(dev))
645 		return dev->limits.max_sg;
646 
647 	/*
648 	 * SRQ allocations are based on powers of 2 for Tavor,
649 	 * (although they only need to be multiples of 16 bytes).
650 	 *
651 	 * Therefore, we need to base the max number of sg entries on
652 	 * the largest power of 2 descriptor size that is <= to the
653 	 * actual max WQE descriptor size, rather than return the
654 	 * max_sg value given by the firmware (which is based on WQE
655 	 * sizes as multiples of 16, not powers of 2).
656 	 *
657 	 * If SRQ implementation is changed for Tavor to be based on
658 	 * multiples of 16, the calculation below can be deleted and
659 	 * the FW max_sg value returned.
660 	 */
661 	return min_t(int, dev->limits.max_sg,
662 		     ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
663 		      sizeof (struct mthca_next_seg)) /
664 		     sizeof (struct mthca_data_seg));
665 }
666 
mthca_init_srq_table(struct mthca_dev * dev)667 int mthca_init_srq_table(struct mthca_dev *dev)
668 {
669 	int err;
670 
671 	if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
672 		return 0;
673 
674 	spin_lock_init(&dev->srq_table.lock);
675 
676 	err = mthca_alloc_init(&dev->srq_table.alloc,
677 			       dev->limits.num_srqs,
678 			       dev->limits.num_srqs - 1,
679 			       dev->limits.reserved_srqs);
680 	if (err)
681 		return err;
682 
683 	err = mthca_array_init(&dev->srq_table.srq,
684 			       dev->limits.num_srqs);
685 	if (err)
686 		mthca_alloc_cleanup(&dev->srq_table.alloc);
687 
688 	return err;
689 }
690 
mthca_cleanup_srq_table(struct mthca_dev * dev)691 void mthca_cleanup_srq_table(struct mthca_dev *dev)
692 {
693 	if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
694 		return;
695 
696 	mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
697 	mthca_alloc_cleanup(&dev->srq_table.alloc);
698 }
699