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);
121 logsize = ilog2(srq->max);
122 context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
123 context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
124 context->db_index = cpu_to_be32(srq->db_index);
125 context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
126 if (pd->ibpd.uobject)
127 context->logstride_usrpage |=
128 cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
130 context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
131 context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
134 static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
136 mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
137 srq->is_direct, &srq->mr);
141 static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
142 struct mthca_srq *srq)
144 struct mthca_data_seg *scatter;
149 if (pd->ibpd.uobject)
152 srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
156 err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
157 MTHCA_MAX_DIRECT_SRQ_SIZE,
158 &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
165 * Now initialize the SRQ buffer so that all of the WQEs are
166 * linked into the list of free WQEs. In addition, set the
167 * scatter list L_Keys to the sentry value of 0x100.
169 for (i = 0; i < srq->max; ++i) {
170 struct mthca_next_seg *next;
172 next = wqe = get_wqe(srq, i);
174 if (i < srq->max - 1) {
175 *wqe_to_link(wqe) = i + 1;
176 next->nda_op = htonl(((i + 1) << srq->wqe_shift) | 1);
178 *wqe_to_link(wqe) = -1;
182 for (scatter = wqe + sizeof (struct mthca_next_seg);
183 (void *) scatter < wqe + (1 << srq->wqe_shift);
185 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
188 srq->last = get_wqe(srq, srq->max - 1);
193 int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
194 struct ib_srq_attr *attr, struct mthca_srq *srq)
196 struct mthca_mailbox *mailbox;
200 /* Sanity check SRQ size before proceeding */
201 if (attr->max_wr > dev->limits.max_srq_wqes ||
202 attr->max_sge > dev->limits.max_srq_sge)
205 srq->max = attr->max_wr;
206 srq->max_gs = attr->max_sge;
209 if (mthca_is_memfree(dev))
210 srq->max = roundup_pow_of_two(srq->max + 1);
212 srq->max = srq->max + 1;
215 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
216 srq->max_gs * sizeof (struct mthca_data_seg)));
218 if (!mthca_is_memfree(dev) && (ds > dev->limits.max_desc_sz))
221 srq->wqe_shift = ilog2(ds);
223 srq->srqn = mthca_alloc(&dev->srq_table.alloc);
227 if (mthca_is_memfree(dev)) {
228 err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
232 if (!pd->ibpd.uobject) {
233 srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
234 srq->srqn, &srq->db);
235 if (srq->db_index < 0) {
242 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
243 if (IS_ERR(mailbox)) {
244 err = PTR_ERR(mailbox);
248 err = mthca_alloc_srq_buf(dev, pd, srq);
250 goto err_out_mailbox;
252 spin_lock_init(&srq->lock);
254 init_waitqueue_head(&srq->wait);
255 mutex_init(&srq->mutex);
257 if (mthca_is_memfree(dev))
258 mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
260 mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
262 err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn);
265 mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
266 goto err_out_free_buf;
269 spin_lock_irq(&dev->srq_table.lock);
270 if (mthca_array_set(&dev->srq_table.srq,
271 srq->srqn & (dev->limits.num_srqs - 1),
273 spin_unlock_irq(&dev->srq_table.lock);
274 goto err_out_free_srq;
276 spin_unlock_irq(&dev->srq_table.lock);
278 mthca_free_mailbox(dev, mailbox);
281 srq->last_free = srq->max - 1;
283 attr->max_wr = srq->max - 1;
284 attr->max_sge = srq->max_gs;
289 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
291 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
294 if (!pd->ibpd.uobject)
295 mthca_free_srq_buf(dev, srq);
298 mthca_free_mailbox(dev, mailbox);
301 if (!pd->ibpd.uobject && mthca_is_memfree(dev))
302 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
305 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
308 mthca_free(&dev->srq_table.alloc, srq->srqn);
313 static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
317 spin_lock_irq(&dev->srq_table.lock);
319 spin_unlock_irq(&dev->srq_table.lock);
324 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
326 struct mthca_mailbox *mailbox;
329 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
330 if (IS_ERR(mailbox)) {
331 mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
335 err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
337 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
339 spin_lock_irq(&dev->srq_table.lock);
340 mthca_array_clear(&dev->srq_table.srq,
341 srq->srqn & (dev->limits.num_srqs - 1));
343 spin_unlock_irq(&dev->srq_table.lock);
345 wait_event(srq->wait, !get_srq_refcount(dev, srq));
347 if (!srq->ibsrq.uobject) {
348 mthca_free_srq_buf(dev, srq);
349 if (mthca_is_memfree(dev))
350 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
353 mthca_table_put(dev, dev->srq_table.table, srq->srqn);
354 mthca_free(&dev->srq_table.alloc, srq->srqn);
355 mthca_free_mailbox(dev, mailbox);
358 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
359 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
361 struct mthca_dev *dev = to_mdev(ibsrq->device);
362 struct mthca_srq *srq = to_msrq(ibsrq);
365 /* We don't support resizing SRQs (yet?) */
366 if (attr_mask & IB_SRQ_MAX_WR)
369 if (attr_mask & IB_SRQ_LIMIT) {
370 u32 max_wr = mthca_is_memfree(dev) ? srq->max - 1 : srq->max;
371 if (attr->srq_limit > max_wr)
374 mutex_lock(&srq->mutex);
375 ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit);
376 mutex_unlock(&srq->mutex);
382 int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
384 struct mthca_dev *dev = to_mdev(ibsrq->device);
385 struct mthca_srq *srq = to_msrq(ibsrq);
386 struct mthca_mailbox *mailbox;
387 struct mthca_arbel_srq_context *arbel_ctx;
388 struct mthca_tavor_srq_context *tavor_ctx;
391 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
393 return PTR_ERR(mailbox);
395 err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox);
399 if (mthca_is_memfree(dev)) {
400 arbel_ctx = mailbox->buf;
401 srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
403 tavor_ctx = mailbox->buf;
404 srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
407 srq_attr->max_wr = srq->max - 1;
408 srq_attr->max_sge = srq->max_gs;
411 mthca_free_mailbox(dev, mailbox);
416 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
417 enum ib_event_type event_type)
419 struct mthca_srq *srq;
420 struct ib_event event;
422 spin_lock(&dev->srq_table.lock);
423 srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
426 spin_unlock(&dev->srq_table.lock);
429 mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
433 if (!srq->ibsrq.event_handler)
436 event.device = &dev->ib_dev;
437 event.event = event_type;
438 event.element.srq = &srq->ibsrq;
439 srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
442 spin_lock(&dev->srq_table.lock);
443 if (!--srq->refcount)
445 spin_unlock(&dev->srq_table.lock);
449 * This function must be called with IRQs disabled.
451 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
454 struct mthca_next_seg *last_free;
456 ind = wqe_addr >> srq->wqe_shift;
458 spin_lock(&srq->lock);
460 last_free = get_wqe(srq, srq->last_free);
461 *wqe_to_link(last_free) = ind;
462 last_free->nda_op = htonl((ind << srq->wqe_shift) | 1);
463 *wqe_to_link(get_wqe(srq, ind)) = -1;
464 srq->last_free = ind;
466 spin_unlock(&srq->lock);
469 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
470 struct ib_recv_wr **bad_wr)
472 struct mthca_dev *dev = to_mdev(ibsrq->device);
473 struct mthca_srq *srq = to_msrq(ibsrq);
484 spin_lock_irqsave(&srq->lock, flags);
486 first_ind = srq->first_free;
488 for (nreq = 0; wr; wr = wr->next) {
489 ind = srq->first_free;
490 wqe = get_wqe(srq, ind);
491 next_ind = *wqe_to_link(wqe);
493 if (unlikely(next_ind < 0)) {
494 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
500 prev_wqe = srq->last;
503 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
504 /* flags field will always remain 0 */
506 wqe += sizeof (struct mthca_next_seg);
508 if (unlikely(wr->num_sge > srq->max_gs)) {
511 srq->last = prev_wqe;
515 for (i = 0; i < wr->num_sge; ++i) {
516 mthca_set_data_seg(wqe, wr->sg_list + i);
517 wqe += sizeof (struct mthca_data_seg);
521 mthca_set_data_seg_inval(wqe);
523 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
524 cpu_to_be32(MTHCA_NEXT_DBD);
526 srq->wrid[ind] = wr->wr_id;
527 srq->first_free = next_ind;
530 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
534 * Make sure that descriptors are written
535 * before doorbell is rung.
539 mthca_write64(first_ind << srq->wqe_shift, srq->srqn << 8,
540 dev->kar + MTHCA_RECEIVE_DOORBELL,
541 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
543 first_ind = srq->first_free;
549 * Make sure that descriptors are written before
554 mthca_write64(first_ind << srq->wqe_shift, (srq->srqn << 8) | nreq,
555 dev->kar + MTHCA_RECEIVE_DOORBELL,
556 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
560 * Make sure doorbells don't leak out of SRQ spinlock and
561 * reach the HCA out of order:
565 spin_unlock_irqrestore(&srq->lock, flags);
569 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
570 struct ib_recv_wr **bad_wr)
572 struct mthca_dev *dev = to_mdev(ibsrq->device);
573 struct mthca_srq *srq = to_msrq(ibsrq);
582 spin_lock_irqsave(&srq->lock, flags);
584 for (nreq = 0; wr; ++nreq, wr = wr->next) {
585 ind = srq->first_free;
586 wqe = get_wqe(srq, ind);
587 next_ind = *wqe_to_link(wqe);
589 if (unlikely(next_ind < 0)) {
590 mthca_err(dev, "SRQ %06x full\n", srq->srqn);
596 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
597 /* flags field will always remain 0 */
599 wqe += sizeof (struct mthca_next_seg);
601 if (unlikely(wr->num_sge > srq->max_gs)) {
607 for (i = 0; i < wr->num_sge; ++i) {
608 mthca_set_data_seg(wqe, wr->sg_list + i);
609 wqe += sizeof (struct mthca_data_seg);
613 mthca_set_data_seg_inval(wqe);
615 srq->wrid[ind] = wr->wr_id;
616 srq->first_free = next_ind;
620 srq->counter += nreq;
623 * Make sure that descriptors are written before
624 * we write doorbell record.
627 *srq->db = cpu_to_be32(srq->counter);
630 spin_unlock_irqrestore(&srq->lock, flags);
634 int mthca_max_srq_sge(struct mthca_dev *dev)
636 if (mthca_is_memfree(dev))
637 return dev->limits.max_sg;
640 * SRQ allocations are based on powers of 2 for Tavor,
641 * (although they only need to be multiples of 16 bytes).
643 * Therefore, we need to base the max number of sg entries on
644 * the largest power of 2 descriptor size that is <= to the
645 * actual max WQE descriptor size, rather than return the
646 * max_sg value given by the firmware (which is based on WQE
647 * sizes as multiples of 16, not powers of 2).
649 * If SRQ implementation is changed for Tavor to be based on
650 * multiples of 16, the calculation below can be deleted and
651 * the FW max_sg value returned.
653 return min_t(int, dev->limits.max_sg,
654 ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
655 sizeof (struct mthca_next_seg)) /
656 sizeof (struct mthca_data_seg));
659 int mthca_init_srq_table(struct mthca_dev *dev)
663 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
666 spin_lock_init(&dev->srq_table.lock);
668 err = mthca_alloc_init(&dev->srq_table.alloc,
669 dev->limits.num_srqs,
670 dev->limits.num_srqs - 1,
671 dev->limits.reserved_srqs);
675 err = mthca_array_init(&dev->srq_table.srq,
676 dev->limits.num_srqs);
678 mthca_alloc_cleanup(&dev->srq_table.alloc);
683 void mthca_cleanup_srq_table(struct mthca_dev *dev)
685 if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
688 mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
689 mthca_alloc_cleanup(&dev->srq_table.alloc);