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
33 #include <linux/slab.h>
34 #include <linux/string.h>
35 #include <linux/sched.h>
39 #include "mthca_dev.h"
40 #include "mthca_cmd.h"
41 #include "mthca_memfree.h"
42 #include "mthca_wqe.h"
45 MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
48 struct mthca_tavor_srq_context {
49 __be64 wqe_base_ds; /* low 6 bits is descriptor size */
53 __be16 limit_watermark;
58 struct mthca_arbel_srq_context {
59 __be32 state_logsize_srqn;
62 __be32 logstride_usrpage;
65 __be16 limit_watermark;
72 static void *get_wqe(struct mthca_srq *srq, int n)
75 return srq->queue.direct.buf + (n << srq->wqe_shift);
77 return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
78 ((n << srq->wqe_shift) & (PAGE_SIZE - 1));
82 * Return a pointer to the location within a WQE that we're using as a
83 * link when the WQE is in the free list. We use the imm field
84 * because in the Tavor case, posting a WQE may overwrite the next
85 * segment of the previous WQE, but a receive WQE will never touch the
86 * imm field. This avoids corrupting our free list if the previous
87 * WQE has already completed and been put on the free list when we
90 static inline int *wqe_to_link(void *wqe)
92 return (int *) (wqe + offsetof(struct mthca_next_seg, imm));
95 static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
97 struct mthca_srq *srq,
98 struct mthca_tavor_srq_context *context)
100 memset(context, 0, sizeof *context);
102 context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
103 context->state_pd = cpu_to_be32(pd->pd_num);
104 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
106 if (pd->ibpd.uobject)
108 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
110 context->uar = cpu_to_be32(dev->driver_uar.index);
113 static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
115 struct mthca_srq *srq,
116 struct mthca_arbel_srq_context *context)
120 memset(context, 0, sizeof *context);
123 * Put max in a temporary variable to work around gcc bug
124 * triggered by ilog2() on sparc64.
127 logsize = ilog2(max);
128 context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
129 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
130 context->db_index = cpu_to_be32(srq->db_index);
131 context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
132 if (pd->ibpd.uobject)
133 context->logstride_usrpage |=
134 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
136 context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
137 context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
140 static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
142 mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
143 srq->is_direct, &srq->mr);
147 static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
148 struct mthca_srq *srq)
150 struct mthca_data_seg *scatter;
155 if (pd->ibpd.uobject)
158 srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
162 err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
163 MTHCA_MAX_DIRECT_SRQ_SIZE,
164 &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
171 * Now initialize the SRQ buffer so that all of the WQEs are
172 * linked into the list of free WQEs. In addition, set the
173 * scatter list L_Keys to the sentry value of 0x100.
175 for (i = 0; i < srq->max; ++i) {
176 struct mthca_next_seg *next;
178 next = wqe = get_wqe(srq, i);
180 if (i < srq->max - 1) {
181 *wqe_to_link(wqe) = i + 1;
182 next->nda_op = htonl(((i + 1) << srq->wqe_shift) | 1);
184 *wqe_to_link(wqe) = -1;
188 for (scatter = wqe + sizeof (struct mthca_next_seg);
189 (void *) scatter < wqe + (1 << srq->wqe_shift);
191 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
194 srq->last = get_wqe(srq, srq->max - 1);
199 int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
200 struct ib_srq_attr *attr, struct mthca_srq *srq)
202 struct mthca_mailbox *mailbox;
206 /* Sanity check SRQ size before proceeding */
207 if (attr->max_wr > dev->limits.max_srq_wqes ||
208 attr->max_sge > dev->limits.max_srq_sge)
211 srq->max = attr->max_wr;
212 srq->max_gs = attr->max_sge;
215 if (mthca_is_memfree(dev))
216 srq->max = roundup_pow_of_two(srq->max + 1);
218 srq->max = srq->max + 1;
221 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
222 srq->max_gs * sizeof (struct mthca_data_seg)));
224 if (!mthca_is_memfree(dev) && (ds > dev->limits.max_desc_sz))
227 srq->wqe_shift = ilog2(ds);
229 srq->srqn = mthca_alloc(&dev->srq_table.alloc);
233 if (mthca_is_memfree(dev)) {
234 err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
238 if (!pd->ibpd.uobject) {
239 srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
240 srq->srqn, &srq->db);
241 if (srq->db_index < 0) {
248 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
249 if (IS_ERR(mailbox)) {
250 err = PTR_ERR(mailbox);
254 err = mthca_alloc_srq_buf(dev, pd, srq);
256 goto err_out_mailbox;
258 spin_lock_init(&srq->lock);
260 init_waitqueue_head(&srq->wait);
261 mutex_init(&srq->mutex);
263 if (mthca_is_memfree(dev))
264 mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
266 mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
268 err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn);
271 mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
272 goto err_out_free_buf;
275 spin_lock_irq(&dev->srq_table.lock);
276 if (mthca_array_set(&dev->srq_table.srq,
277 srq->srqn & (dev->limits.num_srqs - 1),
279 spin_unlock_irq(&dev->srq_table.lock);
280 goto err_out_free_srq;
282 spin_unlock_irq(&dev->srq_table.lock);
284 mthca_free_mailbox(dev, mailbox);
287 srq->last_free = srq->max - 1;
289 attr->max_wr = srq->max - 1;
290 attr->max_sge = srq->max_gs;
295 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
297 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
300 if (!pd->ibpd.uobject)
301 mthca_free_srq_buf(dev, srq);
304 mthca_free_mailbox(dev, mailbox);
307 if (!pd->ibpd.uobject && mthca_is_memfree(dev))
308 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
311 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
314 mthca_free(&dev->srq_table.alloc, srq->srqn);
319 static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
323 spin_lock_irq(&dev->srq_table.lock);
325 spin_unlock_irq(&dev->srq_table.lock);
330 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
332 struct mthca_mailbox *mailbox;
335 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
336 if (IS_ERR(mailbox)) {
337 mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
341 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
343 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
345 spin_lock_irq(&dev->srq_table.lock);
346 mthca_array_clear(&dev->srq_table.srq,
347 srq->srqn & (dev->limits.num_srqs - 1));
349 spin_unlock_irq(&dev->srq_table.lock);
351 wait_event(srq->wait, !get_srq_refcount(dev, srq));
353 if (!srq->ibsrq.uobject) {
354 mthca_free_srq_buf(dev, srq);
355 if (mthca_is_memfree(dev))
356 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
359 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
360 mthca_free(&dev->srq_table.alloc, srq->srqn);
361 mthca_free_mailbox(dev, mailbox);
364 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
365 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
367 struct mthca_dev *dev = to_mdev(ibsrq->device);
368 struct mthca_srq *srq = to_msrq(ibsrq);
371 /* We don't support resizing SRQs (yet?) */
372 if (attr_mask & IB_SRQ_MAX_WR)
375 if (attr_mask & IB_SRQ_LIMIT) {
376 u32 max_wr = mthca_is_memfree(dev) ? srq->max - 1 : srq->max;
377 if (attr->srq_limit > max_wr)
380 mutex_lock(&srq->mutex);
381 ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit);
382 mutex_unlock(&srq->mutex);
388 int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
390 struct mthca_dev *dev = to_mdev(ibsrq->device);
391 struct mthca_srq *srq = to_msrq(ibsrq);
392 struct mthca_mailbox *mailbox;
393 struct mthca_arbel_srq_context *arbel_ctx;
394 struct mthca_tavor_srq_context *tavor_ctx;
397 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
399 return PTR_ERR(mailbox);
401 err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox);
405 if (mthca_is_memfree(dev)) {
406 arbel_ctx = mailbox->buf;
407 srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
409 tavor_ctx = mailbox->buf;
410 srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
413 srq_attr->max_wr = srq->max - 1;
414 srq_attr->max_sge = srq->max_gs;
417 mthca_free_mailbox(dev, mailbox);
422 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
423 enum ib_event_type event_type)
425 struct mthca_srq *srq;
426 struct ib_event event;
428 spin_lock(&dev->srq_table.lock);
429 srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
432 spin_unlock(&dev->srq_table.lock);
435 mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
439 if (!srq->ibsrq.event_handler)
442 event.device = &dev->ib_dev;
443 event.event = event_type;
444 event.element.srq = &srq->ibsrq;
445 srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
448 spin_lock(&dev->srq_table.lock);
449 if (!--srq->refcount)
451 spin_unlock(&dev->srq_table.lock);
455 * This function must be called with IRQs disabled.
457 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
460 struct mthca_next_seg *last_free;
462 ind = wqe_addr >> srq->wqe_shift;
464 spin_lock(&srq->lock);
466 last_free = get_wqe(srq, srq->last_free);
467 *wqe_to_link(last_free) = ind;
468 last_free->nda_op = htonl((ind << srq->wqe_shift) | 1);
469 *wqe_to_link(get_wqe(srq, ind)) = -1;
470 srq->last_free = ind;
472 spin_unlock(&srq->lock);
475 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
476 struct ib_recv_wr **bad_wr)
478 struct mthca_dev *dev = to_mdev(ibsrq->device);
479 struct mthca_srq *srq = to_msrq(ibsrq);
490 spin_lock_irqsave(&srq->lock, flags);
492 first_ind = srq->first_free;
494 for (nreq = 0; wr; wr = wr->next) {
495 ind = srq->first_free;
496 wqe = get_wqe(srq, ind);
497 next_ind = *wqe_to_link(wqe);
499 if (unlikely(next_ind < 0)) {
500 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
506 prev_wqe = srq->last;
509 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
510 /* flags field will always remain 0 */
512 wqe += sizeof (struct mthca_next_seg);
514 if (unlikely(wr->num_sge > srq->max_gs)) {
517 srq->last = prev_wqe;
521 for (i = 0; i < wr->num_sge; ++i) {
522 mthca_set_data_seg(wqe, wr->sg_list + i);
523 wqe += sizeof (struct mthca_data_seg);
527 mthca_set_data_seg_inval(wqe);
529 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
530 cpu_to_be32(MTHCA_NEXT_DBD);
532 srq->wrid[ind] = wr->wr_id;
533 srq->first_free = next_ind;
536 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
540 * Make sure that descriptors are written
541 * before doorbell is rung.
545 mthca_write64(first_ind << srq->wqe_shift, srq->srqn << 8,
546 dev->kar + MTHCA_RECEIVE_DOORBELL,
547 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
549 first_ind = srq->first_free;
555 * Make sure that descriptors are written before
560 mthca_write64(first_ind << srq->wqe_shift, (srq->srqn << 8) | nreq,
561 dev->kar + MTHCA_RECEIVE_DOORBELL,
562 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
566 * Make sure doorbells don't leak out of SRQ spinlock and
567 * reach the HCA out of order:
571 spin_unlock_irqrestore(&srq->lock, flags);
575 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
576 struct ib_recv_wr **bad_wr)
578 struct mthca_dev *dev = to_mdev(ibsrq->device);
579 struct mthca_srq *srq = to_msrq(ibsrq);
588 spin_lock_irqsave(&srq->lock, flags);
590 for (nreq = 0; wr; ++nreq, wr = wr->next) {
591 ind = srq->first_free;
592 wqe = get_wqe(srq, ind);
593 next_ind = *wqe_to_link(wqe);
595 if (unlikely(next_ind < 0)) {
596 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
602 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
603 /* flags field will always remain 0 */
605 wqe += sizeof (struct mthca_next_seg);
607 if (unlikely(wr->num_sge > srq->max_gs)) {
613 for (i = 0; i < wr->num_sge; ++i) {
614 mthca_set_data_seg(wqe, wr->sg_list + i);
615 wqe += sizeof (struct mthca_data_seg);
619 mthca_set_data_seg_inval(wqe);
621 srq->wrid[ind] = wr->wr_id;
622 srq->first_free = next_ind;
626 srq->counter += nreq;
629 * Make sure that descriptors are written before
630 * we write doorbell record.
633 *srq->db = cpu_to_be32(srq->counter);
636 spin_unlock_irqrestore(&srq->lock, flags);
640 int mthca_max_srq_sge(struct mthca_dev *dev)
642 if (mthca_is_memfree(dev))
643 return dev->limits.max_sg;
646 * SRQ allocations are based on powers of 2 for Tavor,
647 * (although they only need to be multiples of 16 bytes).
649 * Therefore, we need to base the max number of sg entries on
650 * the largest power of 2 descriptor size that is <= to the
651 * actual max WQE descriptor size, rather than return the
652 * max_sg value given by the firmware (which is based on WQE
653 * sizes as multiples of 16, not powers of 2).
655 * If SRQ implementation is changed for Tavor to be based on
656 * multiples of 16, the calculation below can be deleted and
657 * the FW max_sg value returned.
659 return min_t(int, dev->limits.max_sg,
660 ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
661 sizeof (struct mthca_next_seg)) /
662 sizeof (struct mthca_data_seg));
665 int mthca_init_srq_table(struct mthca_dev *dev)
669 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
672 spin_lock_init(&dev->srq_table.lock);
674 err = mthca_alloc_init(&dev->srq_table.alloc,
675 dev->limits.num_srqs,
676 dev->limits.num_srqs - 1,
677 dev->limits.reserved_srqs);
681 err = mthca_array_init(&dev->srq_table.srq,
682 dev->limits.num_srqs);
684 mthca_alloc_cleanup(&dev->srq_table.alloc);
689 void mthca_cleanup_srq_table(struct mthca_dev *dev)
691 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
694 mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
695 mthca_alloc_cleanup(&dev->srq_table.alloc);