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 an offset of 4
79 * because in the Tavor case, posting a WQE may overwrite the first
80 * four bytes of the previous WQE. The offset avoids corrupting our
81 * free list if the WQE has already completed and been put on the free
82 * list when we post the next WQE.
84 static inline int *wqe_to_link(void *wqe)
86 return (int *) (wqe + 4);
89 static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
91 struct mthca_srq *srq,
92 struct mthca_tavor_srq_context *context)
94 memset(context, 0, sizeof *context);
96 context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
97 context->state_pd = cpu_to_be32(pd->pd_num);
98 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
100 if (pd->ibpd.uobject)
102 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
104 context->uar = cpu_to_be32(dev->driver_uar.index);
107 static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
109 struct mthca_srq *srq,
110 struct mthca_arbel_srq_context *context)
114 memset(context, 0, sizeof *context);
116 logsize = long_log2(srq->max) + srq->wqe_shift;
117 context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
118 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
119 context->db_index = cpu_to_be32(srq->db_index);
120 context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
121 if (pd->ibpd.uobject)
122 context->logstride_usrpage |=
123 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
125 context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
126 context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
129 static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
131 mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
132 srq->is_direct, &srq->mr);
136 static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
137 struct mthca_srq *srq)
139 struct mthca_data_seg *scatter;
144 if (pd->ibpd.uobject)
147 srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
151 err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
152 MTHCA_MAX_DIRECT_SRQ_SIZE,
153 &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
160 * Now initialize the SRQ buffer so that all of the WQEs are
161 * linked into the list of free WQEs. In addition, set the
162 * scatter list L_Keys to the sentry value of 0x100.
164 for (i = 0; i < srq->max; ++i) {
165 wqe = get_wqe(srq, i);
167 *wqe_to_link(wqe) = i < srq->max - 1 ? i + 1 : -1;
169 for (scatter = wqe + sizeof (struct mthca_next_seg);
170 (void *) scatter < wqe + (1 << srq->wqe_shift);
172 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
178 int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
179 struct ib_srq_attr *attr, struct mthca_srq *srq)
181 struct mthca_mailbox *mailbox;
186 /* Sanity check SRQ size before proceeding */
187 if (attr->max_wr > 16 << 20 || attr->max_sge > 64)
190 srq->max = attr->max_wr;
191 srq->max_gs = attr->max_sge;
195 if (mthca_is_memfree(dev))
196 srq->max = roundup_pow_of_two(srq->max + 1);
199 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
200 srq->max_gs * sizeof (struct mthca_data_seg)));
201 srq->wqe_shift = long_log2(ds);
203 srq->srqn = mthca_alloc(&dev->srq_table.alloc);
207 if (mthca_is_memfree(dev)) {
208 err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
212 if (!pd->ibpd.uobject) {
213 srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
214 srq->srqn, &srq->db);
215 if (srq->db_index < 0) {
222 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
223 if (IS_ERR(mailbox)) {
224 err = PTR_ERR(mailbox);
228 err = mthca_alloc_srq_buf(dev, pd, srq);
230 goto err_out_mailbox;
232 spin_lock_init(&srq->lock);
233 atomic_set(&srq->refcount, 1);
234 init_waitqueue_head(&srq->wait);
236 if (mthca_is_memfree(dev))
237 mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
239 mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
241 err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn, &status);
244 mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
245 goto err_out_free_buf;
248 mthca_warn(dev, "SW2HW_SRQ returned status 0x%02x\n",
251 goto err_out_free_buf;
254 spin_lock_irq(&dev->srq_table.lock);
255 if (mthca_array_set(&dev->srq_table.srq,
256 srq->srqn & (dev->limits.num_srqs - 1),
258 spin_unlock_irq(&dev->srq_table.lock);
259 goto err_out_free_srq;
261 spin_unlock_irq(&dev->srq_table.lock);
263 mthca_free_mailbox(dev, mailbox);
266 srq->last_free = srq->max - 1;
271 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
273 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
275 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
278 if (!pd->ibpd.uobject)
279 mthca_free_srq_buf(dev, srq);
282 mthca_free_mailbox(dev, mailbox);
285 if (!pd->ibpd.uobject && mthca_is_memfree(dev))
286 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
289 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
292 mthca_free(&dev->srq_table.alloc, srq->srqn);
297 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
299 struct mthca_mailbox *mailbox;
303 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
304 if (IS_ERR(mailbox)) {
305 mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
309 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
311 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
313 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
315 spin_lock_irq(&dev->srq_table.lock);
316 mthca_array_clear(&dev->srq_table.srq,
317 srq->srqn & (dev->limits.num_srqs - 1));
318 spin_unlock_irq(&dev->srq_table.lock);
320 atomic_dec(&srq->refcount);
321 wait_event(srq->wait, !atomic_read(&srq->refcount));
323 if (!srq->ibsrq.uobject) {
324 mthca_free_srq_buf(dev, srq);
325 if (mthca_is_memfree(dev))
326 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
329 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
330 mthca_free(&dev->srq_table.alloc, srq->srqn);
331 mthca_free_mailbox(dev, mailbox);
334 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
335 enum ib_event_type event_type)
337 struct mthca_srq *srq;
338 struct ib_event event;
340 spin_lock(&dev->srq_table.lock);
341 srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
343 atomic_inc(&srq->refcount);
344 spin_unlock(&dev->srq_table.lock);
347 mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
351 if (!srq->ibsrq.event_handler)
354 event.device = &dev->ib_dev;
355 event.event = event_type;
356 event.element.srq = &srq->ibsrq;
357 srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
360 if (atomic_dec_and_test(&srq->refcount))
365 * This function must be called with IRQs disabled.
367 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
371 ind = wqe_addr >> srq->wqe_shift;
373 spin_lock(&srq->lock);
375 if (likely(srq->first_free >= 0))
376 *wqe_to_link(get_wqe(srq, srq->last_free)) = ind;
378 srq->first_free = ind;
380 *wqe_to_link(get_wqe(srq, ind)) = -1;
381 srq->last_free = ind;
383 spin_unlock(&srq->lock);
386 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
387 struct ib_recv_wr **bad_wr)
389 struct mthca_dev *dev = to_mdev(ibsrq->device);
390 struct mthca_srq *srq = to_msrq(ibsrq);
401 spin_lock_irqsave(&srq->lock, flags);
403 first_ind = srq->first_free;
405 for (nreq = 0; wr; ++nreq, wr = wr->next) {
406 ind = srq->first_free;
409 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
415 wqe = get_wqe(srq, ind);
416 next_ind = *wqe_to_link(wqe);
417 prev_wqe = srq->last;
420 ((struct mthca_next_seg *) wqe)->nda_op = 0;
421 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
422 /* flags field will always remain 0 */
424 wqe += sizeof (struct mthca_next_seg);
426 if (unlikely(wr->num_sge > srq->max_gs)) {
429 srq->last = prev_wqe;
433 for (i = 0; i < wr->num_sge; ++i) {
434 ((struct mthca_data_seg *) wqe)->byte_count =
435 cpu_to_be32(wr->sg_list[i].length);
436 ((struct mthca_data_seg *) wqe)->lkey =
437 cpu_to_be32(wr->sg_list[i].lkey);
438 ((struct mthca_data_seg *) wqe)->addr =
439 cpu_to_be64(wr->sg_list[i].addr);
440 wqe += sizeof (struct mthca_data_seg);
443 if (i < srq->max_gs) {
444 ((struct mthca_data_seg *) wqe)->byte_count = 0;
445 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
446 ((struct mthca_data_seg *) wqe)->addr = 0;
449 if (likely(prev_wqe)) {
450 ((struct mthca_next_seg *) prev_wqe)->nda_op =
451 cpu_to_be32((ind << srq->wqe_shift) | 1);
453 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
454 cpu_to_be32(MTHCA_NEXT_DBD);
457 srq->wrid[ind] = wr->wr_id;
458 srq->first_free = next_ind;
466 doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
467 doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
470 * Make sure that descriptors are written before
475 mthca_write64(doorbell,
476 dev->kar + MTHCA_RECEIVE_DOORBELL,
477 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
480 spin_unlock_irqrestore(&srq->lock, flags);
484 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
485 struct ib_recv_wr **bad_wr)
487 struct mthca_dev *dev = to_mdev(ibsrq->device);
488 struct mthca_srq *srq = to_msrq(ibsrq);
497 spin_lock_irqsave(&srq->lock, flags);
499 for (nreq = 0; wr; ++nreq, wr = wr->next) {
500 ind = srq->first_free;
503 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
509 wqe = get_wqe(srq, ind);
510 next_ind = *wqe_to_link(wqe);
512 ((struct mthca_next_seg *) wqe)->nda_op =
513 cpu_to_be32((next_ind << srq->wqe_shift) | 1);
514 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
515 /* flags field will always remain 0 */
517 wqe += sizeof (struct mthca_next_seg);
519 if (unlikely(wr->num_sge > srq->max_gs)) {
525 for (i = 0; i < wr->num_sge; ++i) {
526 ((struct mthca_data_seg *) wqe)->byte_count =
527 cpu_to_be32(wr->sg_list[i].length);
528 ((struct mthca_data_seg *) wqe)->lkey =
529 cpu_to_be32(wr->sg_list[i].lkey);
530 ((struct mthca_data_seg *) wqe)->addr =
531 cpu_to_be64(wr->sg_list[i].addr);
532 wqe += sizeof (struct mthca_data_seg);
535 if (i < srq->max_gs) {
536 ((struct mthca_data_seg *) wqe)->byte_count = 0;
537 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
538 ((struct mthca_data_seg *) wqe)->addr = 0;
541 srq->wrid[ind] = wr->wr_id;
542 srq->first_free = next_ind;
546 srq->counter += nreq;
549 * Make sure that descriptors are written before
550 * we write doorbell record.
553 *srq->db = cpu_to_be32(srq->counter);
556 spin_unlock_irqrestore(&srq->lock, flags);
560 int __devinit mthca_init_srq_table(struct mthca_dev *dev)
564 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
567 spin_lock_init(&dev->srq_table.lock);
569 err = mthca_alloc_init(&dev->srq_table.alloc,
570 dev->limits.num_srqs,
571 dev->limits.num_srqs - 1,
572 dev->limits.reserved_srqs);
576 err = mthca_array_init(&dev->srq_table.srq,
577 dev->limits.num_srqs);
579 mthca_alloc_cleanup(&dev->srq_table.alloc);
584 void __devexit mthca_cleanup_srq_table(struct mthca_dev *dev)
586 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
589 mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
590 mthca_alloc_cleanup(&dev->srq_table.alloc);