2 * Copyright (c) 2005 Cisco Systems. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 * $Id: mthca_srq.c 3047 2005-08-10 03:59:35Z roland $
35 #include "mthca_dev.h"
36 #include "mthca_cmd.h"
37 #include "mthca_memfree.h"
38 #include "mthca_wqe.h"
41 MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
44 struct mthca_tavor_srq_context {
45 __be64 wqe_base_ds; /* low 6 bits is descriptor size */
53 struct mthca_arbel_srq_context {
54 __be32 state_logsize_srqn;
57 __be32 logstride_usrpage;
60 __be16 limit_watermark;
67 static void *get_wqe(struct mthca_srq *srq, int n)
70 return srq->queue.direct.buf + (n << srq->wqe_shift);
72 return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
73 ((n << srq->wqe_shift) & (PAGE_SIZE - 1));
77 * Return a pointer to the location within a WQE that we're using as a
78 * link when the WQE is in the free list. We use the imm field
79 * because in the Tavor case, posting a WQE may overwrite the next
80 * segment of the previous WQE, but a receive WQE will never touch the
81 * imm field. This avoids corrupting our free list if the previous
82 * WQE has already completed and been put on the free list when we
85 static inline int *wqe_to_link(void *wqe)
87 return (int *) (wqe + offsetof(struct mthca_next_seg, imm));
90 static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
92 struct mthca_srq *srq,
93 struct mthca_tavor_srq_context *context)
95 memset(context, 0, sizeof *context);
97 context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
98 context->state_pd = cpu_to_be32(pd->pd_num);
99 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
101 if (pd->ibpd.uobject)
103 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
105 context->uar = cpu_to_be32(dev->driver_uar.index);
108 static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
110 struct mthca_srq *srq,
111 struct mthca_arbel_srq_context *context)
115 memset(context, 0, sizeof *context);
117 logsize = long_log2(srq->max) + srq->wqe_shift;
118 context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
119 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
120 context->db_index = cpu_to_be32(srq->db_index);
121 context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
122 if (pd->ibpd.uobject)
123 context->logstride_usrpage |=
124 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
126 context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
127 context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
130 static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
132 mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
133 srq->is_direct, &srq->mr);
137 static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
138 struct mthca_srq *srq)
140 struct mthca_data_seg *scatter;
145 if (pd->ibpd.uobject)
148 srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
152 err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
153 MTHCA_MAX_DIRECT_SRQ_SIZE,
154 &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
161 * Now initialize the SRQ buffer so that all of the WQEs are
162 * linked into the list of free WQEs. In addition, set the
163 * scatter list L_Keys to the sentry value of 0x100.
165 for (i = 0; i < srq->max; ++i) {
166 wqe = get_wqe(srq, i);
168 *wqe_to_link(wqe) = i < srq->max - 1 ? i + 1 : -1;
170 for (scatter = wqe + sizeof (struct mthca_next_seg);
171 (void *) scatter < wqe + (1 << srq->wqe_shift);
173 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
176 srq->last = get_wqe(srq, srq->max - 1);
181 int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
182 struct ib_srq_attr *attr, struct mthca_srq *srq)
184 struct mthca_mailbox *mailbox;
189 /* Sanity check SRQ size before proceeding */
190 if (attr->max_wr > dev->limits.max_srq_wqes ||
191 attr->max_sge > dev->limits.max_sg)
194 srq->max = attr->max_wr;
195 srq->max_gs = attr->max_sge;
198 if (mthca_is_memfree(dev))
199 srq->max = roundup_pow_of_two(srq->max + 1);
202 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
203 srq->max_gs * sizeof (struct mthca_data_seg)));
204 srq->wqe_shift = long_log2(ds);
206 srq->srqn = mthca_alloc(&dev->srq_table.alloc);
210 if (mthca_is_memfree(dev)) {
211 err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
215 if (!pd->ibpd.uobject) {
216 srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
217 srq->srqn, &srq->db);
218 if (srq->db_index < 0) {
225 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
226 if (IS_ERR(mailbox)) {
227 err = PTR_ERR(mailbox);
231 err = mthca_alloc_srq_buf(dev, pd, srq);
233 goto err_out_mailbox;
235 spin_lock_init(&srq->lock);
236 atomic_set(&srq->refcount, 1);
237 init_waitqueue_head(&srq->wait);
239 if (mthca_is_memfree(dev))
240 mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
242 mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
244 err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn, &status);
247 mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
248 goto err_out_free_buf;
251 mthca_warn(dev, "SW2HW_SRQ returned status 0x%02x\n",
254 goto err_out_free_buf;
257 spin_lock_irq(&dev->srq_table.lock);
258 if (mthca_array_set(&dev->srq_table.srq,
259 srq->srqn & (dev->limits.num_srqs - 1),
261 spin_unlock_irq(&dev->srq_table.lock);
262 goto err_out_free_srq;
264 spin_unlock_irq(&dev->srq_table.lock);
266 mthca_free_mailbox(dev, mailbox);
269 srq->last_free = srq->max - 1;
274 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
276 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
278 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
281 if (!pd->ibpd.uobject)
282 mthca_free_srq_buf(dev, srq);
285 mthca_free_mailbox(dev, mailbox);
288 if (!pd->ibpd.uobject && mthca_is_memfree(dev))
289 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
292 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
295 mthca_free(&dev->srq_table.alloc, srq->srqn);
300 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
302 struct mthca_mailbox *mailbox;
306 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
307 if (IS_ERR(mailbox)) {
308 mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
312 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
314 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
316 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
318 spin_lock_irq(&dev->srq_table.lock);
319 mthca_array_clear(&dev->srq_table.srq,
320 srq->srqn & (dev->limits.num_srqs - 1));
321 spin_unlock_irq(&dev->srq_table.lock);
323 atomic_dec(&srq->refcount);
324 wait_event(srq->wait, !atomic_read(&srq->refcount));
326 if (!srq->ibsrq.uobject) {
327 mthca_free_srq_buf(dev, srq);
328 if (mthca_is_memfree(dev))
329 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
332 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
333 mthca_free(&dev->srq_table.alloc, srq->srqn);
334 mthca_free_mailbox(dev, mailbox);
337 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
338 enum ib_srq_attr_mask attr_mask)
340 struct mthca_dev *dev = to_mdev(ibsrq->device);
341 struct mthca_srq *srq = to_msrq(ibsrq);
345 /* We don't support resizing SRQs (yet?) */
346 if (attr_mask & IB_SRQ_MAX_WR)
349 if (attr_mask & IB_SRQ_LIMIT) {
350 ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit, &status);
360 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
361 enum ib_event_type event_type)
363 struct mthca_srq *srq;
364 struct ib_event event;
366 spin_lock(&dev->srq_table.lock);
367 srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
369 atomic_inc(&srq->refcount);
370 spin_unlock(&dev->srq_table.lock);
373 mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
377 if (!srq->ibsrq.event_handler)
380 event.device = &dev->ib_dev;
381 event.event = event_type;
382 event.element.srq = &srq->ibsrq;
383 srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
386 if (atomic_dec_and_test(&srq->refcount))
391 * This function must be called with IRQs disabled.
393 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
397 ind = wqe_addr >> srq->wqe_shift;
399 spin_lock(&srq->lock);
401 if (likely(srq->first_free >= 0))
402 *wqe_to_link(get_wqe(srq, srq->last_free)) = ind;
404 srq->first_free = ind;
406 *wqe_to_link(get_wqe(srq, ind)) = -1;
407 srq->last_free = ind;
409 spin_unlock(&srq->lock);
412 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
413 struct ib_recv_wr **bad_wr)
415 struct mthca_dev *dev = to_mdev(ibsrq->device);
416 struct mthca_srq *srq = to_msrq(ibsrq);
427 spin_lock_irqsave(&srq->lock, flags);
429 first_ind = srq->first_free;
431 for (nreq = 0; wr; ++nreq, wr = wr->next) {
432 ind = srq->first_free;
435 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
441 wqe = get_wqe(srq, ind);
442 next_ind = *wqe_to_link(wqe);
445 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
451 prev_wqe = srq->last;
454 ((struct mthca_next_seg *) wqe)->nda_op = 0;
455 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
456 /* flags field will always remain 0 */
458 wqe += sizeof (struct mthca_next_seg);
460 if (unlikely(wr->num_sge > srq->max_gs)) {
463 srq->last = prev_wqe;
467 for (i = 0; i < wr->num_sge; ++i) {
468 ((struct mthca_data_seg *) wqe)->byte_count =
469 cpu_to_be32(wr->sg_list[i].length);
470 ((struct mthca_data_seg *) wqe)->lkey =
471 cpu_to_be32(wr->sg_list[i].lkey);
472 ((struct mthca_data_seg *) wqe)->addr =
473 cpu_to_be64(wr->sg_list[i].addr);
474 wqe += sizeof (struct mthca_data_seg);
477 if (i < srq->max_gs) {
478 ((struct mthca_data_seg *) wqe)->byte_count = 0;
479 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
480 ((struct mthca_data_seg *) wqe)->addr = 0;
483 ((struct mthca_next_seg *) prev_wqe)->nda_op =
484 cpu_to_be32((ind << srq->wqe_shift) | 1);
486 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
487 cpu_to_be32(MTHCA_NEXT_DBD);
489 srq->wrid[ind] = wr->wr_id;
490 srq->first_free = next_ind;
496 doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
497 doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
500 * Make sure that descriptors are written before
505 mthca_write64(doorbell,
506 dev->kar + MTHCA_RECEIVE_DOORBELL,
507 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
510 spin_unlock_irqrestore(&srq->lock, flags);
514 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
515 struct ib_recv_wr **bad_wr)
517 struct mthca_dev *dev = to_mdev(ibsrq->device);
518 struct mthca_srq *srq = to_msrq(ibsrq);
527 spin_lock_irqsave(&srq->lock, flags);
529 for (nreq = 0; wr; ++nreq, wr = wr->next) {
530 ind = srq->first_free;
533 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
539 wqe = get_wqe(srq, ind);
540 next_ind = *wqe_to_link(wqe);
543 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
549 ((struct mthca_next_seg *) wqe)->nda_op =
550 cpu_to_be32((next_ind << srq->wqe_shift) | 1);
551 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
552 /* flags field will always remain 0 */
554 wqe += sizeof (struct mthca_next_seg);
556 if (unlikely(wr->num_sge > srq->max_gs)) {
562 for (i = 0; i < wr->num_sge; ++i) {
563 ((struct mthca_data_seg *) wqe)->byte_count =
564 cpu_to_be32(wr->sg_list[i].length);
565 ((struct mthca_data_seg *) wqe)->lkey =
566 cpu_to_be32(wr->sg_list[i].lkey);
567 ((struct mthca_data_seg *) wqe)->addr =
568 cpu_to_be64(wr->sg_list[i].addr);
569 wqe += sizeof (struct mthca_data_seg);
572 if (i < srq->max_gs) {
573 ((struct mthca_data_seg *) wqe)->byte_count = 0;
574 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
575 ((struct mthca_data_seg *) wqe)->addr = 0;
578 srq->wrid[ind] = wr->wr_id;
579 srq->first_free = next_ind;
583 srq->counter += nreq;
586 * Make sure that descriptors are written before
587 * we write doorbell record.
590 *srq->db = cpu_to_be32(srq->counter);
593 spin_unlock_irqrestore(&srq->lock, flags);
597 int __devinit mthca_init_srq_table(struct mthca_dev *dev)
601 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
604 spin_lock_init(&dev->srq_table.lock);
606 err = mthca_alloc_init(&dev->srq_table.alloc,
607 dev->limits.num_srqs,
608 dev->limits.num_srqs - 1,
609 dev->limits.reserved_srqs);
613 err = mthca_array_init(&dev->srq_table.srq,
614 dev->limits.num_srqs);
616 mthca_alloc_cleanup(&dev->srq_table.alloc);
621 void __devexit mthca_cleanup_srq_table(struct mthca_dev *dev)
623 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
626 mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
627 mthca_alloc_cleanup(&dev->srq_table.alloc);