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 <linux/slab.h>
36 #include <linux/string.h>
38 #include "mthca_dev.h"
39 #include "mthca_cmd.h"
40 #include "mthca_memfree.h"
41 #include "mthca_wqe.h"
44 MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
47 struct mthca_tavor_srq_context {
48 __be64 wqe_base_ds; /* low 6 bits is descriptor size */
56 struct mthca_arbel_srq_context {
57 __be32 state_logsize_srqn;
60 __be32 logstride_usrpage;
63 __be16 limit_watermark;
70 static void *get_wqe(struct mthca_srq *srq, int n)
73 return srq->queue.direct.buf + (n << srq->wqe_shift);
75 return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
76 ((n << srq->wqe_shift) & (PAGE_SIZE - 1));
80 * Return a pointer to the location within a WQE that we're using as a
81 * link when the WQE is in the free list. We use the imm field
82 * because in the Tavor case, posting a WQE may overwrite the next
83 * segment of the previous WQE, but a receive WQE will never touch the
84 * imm field. This avoids corrupting our free list if the previous
85 * WQE has already completed and been put on the free list when we
88 static inline int *wqe_to_link(void *wqe)
90 return (int *) (wqe + offsetof(struct mthca_next_seg, imm));
93 static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
95 struct mthca_srq *srq,
96 struct mthca_tavor_srq_context *context)
98 memset(context, 0, sizeof *context);
100 context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
101 context->state_pd = cpu_to_be32(pd->pd_num);
102 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
104 if (pd->ibpd.uobject)
106 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
108 context->uar = cpu_to_be32(dev->driver_uar.index);
111 static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
113 struct mthca_srq *srq,
114 struct mthca_arbel_srq_context *context)
118 memset(context, 0, sizeof *context);
120 logsize = long_log2(srq->max) + srq->wqe_shift;
121 context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
122 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
123 context->db_index = cpu_to_be32(srq->db_index);
124 context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
125 if (pd->ibpd.uobject)
126 context->logstride_usrpage |=
127 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
129 context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
130 context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
133 static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
135 mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
136 srq->is_direct, &srq->mr);
140 static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
141 struct mthca_srq *srq)
143 struct mthca_data_seg *scatter;
148 if (pd->ibpd.uobject)
151 srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
155 err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
156 MTHCA_MAX_DIRECT_SRQ_SIZE,
157 &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
164 * Now initialize the SRQ buffer so that all of the WQEs are
165 * linked into the list of free WQEs. In addition, set the
166 * scatter list L_Keys to the sentry value of 0x100.
168 for (i = 0; i < srq->max; ++i) {
169 wqe = get_wqe(srq, i);
171 *wqe_to_link(wqe) = i < srq->max - 1 ? i + 1 : -1;
173 for (scatter = wqe + sizeof (struct mthca_next_seg);
174 (void *) scatter < wqe + (1 << srq->wqe_shift);
176 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
179 srq->last = get_wqe(srq, srq->max - 1);
184 int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
185 struct ib_srq_attr *attr, struct mthca_srq *srq)
187 struct mthca_mailbox *mailbox;
192 /* Sanity check SRQ size before proceeding */
193 if (attr->max_wr > dev->limits.max_srq_wqes ||
194 attr->max_sge > dev->limits.max_sg)
197 srq->max = attr->max_wr;
198 srq->max_gs = attr->max_sge;
201 if (mthca_is_memfree(dev))
202 srq->max = roundup_pow_of_two(srq->max + 1);
205 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
206 srq->max_gs * sizeof (struct mthca_data_seg)));
207 srq->wqe_shift = long_log2(ds);
209 srq->srqn = mthca_alloc(&dev->srq_table.alloc);
213 if (mthca_is_memfree(dev)) {
214 err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
218 if (!pd->ibpd.uobject) {
219 srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
220 srq->srqn, &srq->db);
221 if (srq->db_index < 0) {
228 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
229 if (IS_ERR(mailbox)) {
230 err = PTR_ERR(mailbox);
234 err = mthca_alloc_srq_buf(dev, pd, srq);
236 goto err_out_mailbox;
238 spin_lock_init(&srq->lock);
239 atomic_set(&srq->refcount, 1);
240 init_waitqueue_head(&srq->wait);
242 if (mthca_is_memfree(dev))
243 mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
245 mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
247 err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn, &status);
250 mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
251 goto err_out_free_buf;
254 mthca_warn(dev, "SW2HW_SRQ returned status 0x%02x\n",
257 goto err_out_free_buf;
260 spin_lock_irq(&dev->srq_table.lock);
261 if (mthca_array_set(&dev->srq_table.srq,
262 srq->srqn & (dev->limits.num_srqs - 1),
264 spin_unlock_irq(&dev->srq_table.lock);
265 goto err_out_free_srq;
267 spin_unlock_irq(&dev->srq_table.lock);
269 mthca_free_mailbox(dev, mailbox);
272 srq->last_free = srq->max - 1;
274 attr->max_wr = (mthca_is_memfree(dev)) ? srq->max - 1 : srq->max;
275 attr->max_sge = srq->max_gs;
280 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
282 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
284 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
287 if (!pd->ibpd.uobject)
288 mthca_free_srq_buf(dev, srq);
291 mthca_free_mailbox(dev, mailbox);
294 if (!pd->ibpd.uobject && mthca_is_memfree(dev))
295 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
298 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
301 mthca_free(&dev->srq_table.alloc, srq->srqn);
306 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
308 struct mthca_mailbox *mailbox;
312 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
313 if (IS_ERR(mailbox)) {
314 mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
318 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
320 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
322 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
324 spin_lock_irq(&dev->srq_table.lock);
325 mthca_array_clear(&dev->srq_table.srq,
326 srq->srqn & (dev->limits.num_srqs - 1));
327 spin_unlock_irq(&dev->srq_table.lock);
329 atomic_dec(&srq->refcount);
330 wait_event(srq->wait, !atomic_read(&srq->refcount));
332 if (!srq->ibsrq.uobject) {
333 mthca_free_srq_buf(dev, srq);
334 if (mthca_is_memfree(dev))
335 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
338 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
339 mthca_free(&dev->srq_table.alloc, srq->srqn);
340 mthca_free_mailbox(dev, mailbox);
343 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
344 enum ib_srq_attr_mask attr_mask)
346 struct mthca_dev *dev = to_mdev(ibsrq->device);
347 struct mthca_srq *srq = to_msrq(ibsrq);
351 /* We don't support resizing SRQs (yet?) */
352 if (attr_mask & IB_SRQ_MAX_WR)
355 if (attr_mask & IB_SRQ_LIMIT) {
356 ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit, &status);
366 int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
368 struct mthca_dev *dev = to_mdev(ibsrq->device);
369 struct mthca_srq *srq = to_msrq(ibsrq);
370 struct mthca_mailbox *mailbox;
371 struct mthca_arbel_srq_context *arbel_ctx;
375 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
377 return PTR_ERR(mailbox);
379 err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox, &status);
383 if (mthca_is_memfree(dev)) {
384 arbel_ctx = mailbox->buf;
385 srq_attr->srq_limit = arbel_ctx->limit_watermark;
387 srq_attr->srq_limit = 0;
389 srq_attr->max_wr = (mthca_is_memfree(dev)) ? srq->max - 1 : srq->max;
390 srq_attr->max_sge = srq->max_gs;
393 mthca_free_mailbox(dev, mailbox);
398 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
399 enum ib_event_type event_type)
401 struct mthca_srq *srq;
402 struct ib_event event;
404 spin_lock(&dev->srq_table.lock);
405 srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
407 atomic_inc(&srq->refcount);
408 spin_unlock(&dev->srq_table.lock);
411 mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
415 if (!srq->ibsrq.event_handler)
418 event.device = &dev->ib_dev;
419 event.event = event_type;
420 event.element.srq = &srq->ibsrq;
421 srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
424 if (atomic_dec_and_test(&srq->refcount))
429 * This function must be called with IRQs disabled.
431 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
435 ind = wqe_addr >> srq->wqe_shift;
437 spin_lock(&srq->lock);
439 if (likely(srq->first_free >= 0))
440 *wqe_to_link(get_wqe(srq, srq->last_free)) = ind;
442 srq->first_free = ind;
444 *wqe_to_link(get_wqe(srq, ind)) = -1;
445 srq->last_free = ind;
447 spin_unlock(&srq->lock);
450 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
451 struct ib_recv_wr **bad_wr)
453 struct mthca_dev *dev = to_mdev(ibsrq->device);
454 struct mthca_srq *srq = to_msrq(ibsrq);
466 spin_lock_irqsave(&srq->lock, flags);
468 first_ind = srq->first_free;
470 for (nreq = 0; wr; ++nreq, wr = wr->next) {
471 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
474 doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
475 doorbell[1] = cpu_to_be32(srq->srqn << 8);
478 * Make sure that descriptors are written
479 * before doorbell is rung.
483 mthca_write64(doorbell,
484 dev->kar + MTHCA_RECEIVE_DOORBELL,
485 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
487 first_ind = srq->first_free;
490 ind = srq->first_free;
493 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
499 wqe = get_wqe(srq, ind);
500 next_ind = *wqe_to_link(wqe);
503 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
509 prev_wqe = srq->last;
512 ((struct mthca_next_seg *) wqe)->nda_op = 0;
513 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
514 /* flags field will always remain 0 */
516 wqe += sizeof (struct mthca_next_seg);
518 if (unlikely(wr->num_sge > srq->max_gs)) {
521 srq->last = prev_wqe;
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 ((struct mthca_next_seg *) prev_wqe)->nda_op =
542 cpu_to_be32((ind << srq->wqe_shift) | 1);
544 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
545 cpu_to_be32(MTHCA_NEXT_DBD);
547 srq->wrid[ind] = wr->wr_id;
548 srq->first_free = next_ind;
552 doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
553 doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
556 * Make sure that descriptors are written before
561 mthca_write64(doorbell,
562 dev->kar + MTHCA_RECEIVE_DOORBELL,
563 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
566 spin_unlock_irqrestore(&srq->lock, flags);
570 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
571 struct ib_recv_wr **bad_wr)
573 struct mthca_dev *dev = to_mdev(ibsrq->device);
574 struct mthca_srq *srq = to_msrq(ibsrq);
583 spin_lock_irqsave(&srq->lock, flags);
585 for (nreq = 0; wr; ++nreq, wr = wr->next) {
586 ind = srq->first_free;
589 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
595 wqe = get_wqe(srq, ind);
596 next_ind = *wqe_to_link(wqe);
599 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
605 ((struct mthca_next_seg *) wqe)->nda_op =
606 cpu_to_be32((next_ind << srq->wqe_shift) | 1);
607 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
608 /* flags field will always remain 0 */
610 wqe += sizeof (struct mthca_next_seg);
612 if (unlikely(wr->num_sge > srq->max_gs)) {
618 for (i = 0; i < wr->num_sge; ++i) {
619 ((struct mthca_data_seg *) wqe)->byte_count =
620 cpu_to_be32(wr->sg_list[i].length);
621 ((struct mthca_data_seg *) wqe)->lkey =
622 cpu_to_be32(wr->sg_list[i].lkey);
623 ((struct mthca_data_seg *) wqe)->addr =
624 cpu_to_be64(wr->sg_list[i].addr);
625 wqe += sizeof (struct mthca_data_seg);
628 if (i < srq->max_gs) {
629 ((struct mthca_data_seg *) wqe)->byte_count = 0;
630 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
631 ((struct mthca_data_seg *) wqe)->addr = 0;
634 srq->wrid[ind] = wr->wr_id;
635 srq->first_free = next_ind;
639 srq->counter += nreq;
642 * Make sure that descriptors are written before
643 * we write doorbell record.
646 *srq->db = cpu_to_be32(srq->counter);
649 spin_unlock_irqrestore(&srq->lock, flags);
653 int __devinit mthca_init_srq_table(struct mthca_dev *dev)
657 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
660 spin_lock_init(&dev->srq_table.lock);
662 err = mthca_alloc_init(&dev->srq_table.alloc,
663 dev->limits.num_srqs,
664 dev->limits.num_srqs - 1,
665 dev->limits.reserved_srqs);
669 err = mthca_array_init(&dev->srq_table.srq,
670 dev->limits.num_srqs);
672 mthca_alloc_cleanup(&dev->srq_table.alloc);
677 void __devexit mthca_cleanup_srq_table(struct mthca_dev *dev)
679 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
682 mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
683 mthca_alloc_cleanup(&dev->srq_table.alloc);