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
;
207 /* Sanity check SRQ size before proceeding */
208 if (attr
->max_wr
> dev
->limits
.max_srq_wqes
||
209 attr
->max_sge
> dev
->limits
.max_srq_sge
)
212 srq
->max
= attr
->max_wr
;
213 srq
->max_gs
= attr
->max_sge
;
216 if (mthca_is_memfree(dev
))
217 srq
->max
= roundup_pow_of_two(srq
->max
+ 1);
219 srq
->max
= srq
->max
+ 1;
222 roundup_pow_of_two(sizeof (struct mthca_next_seg
) +
223 srq
->max_gs
* sizeof (struct mthca_data_seg
)));
225 if (!mthca_is_memfree(dev
) && (ds
> dev
->limits
.max_desc_sz
))
228 srq
->wqe_shift
= ilog2(ds
);
230 srq
->srqn
= mthca_alloc(&dev
->srq_table
.alloc
);
234 if (mthca_is_memfree(dev
)) {
235 err
= mthca_table_get(dev
, dev
->srq_table
.table
, srq
->srqn
);
239 if (!pd
->ibpd
.uobject
) {
240 srq
->db_index
= mthca_alloc_db(dev
, MTHCA_DB_TYPE_SRQ
,
241 srq
->srqn
, &srq
->db
);
242 if (srq
->db_index
< 0) {
249 mailbox
= mthca_alloc_mailbox(dev
, GFP_KERNEL
);
250 if (IS_ERR(mailbox
)) {
251 err
= PTR_ERR(mailbox
);
255 err
= mthca_alloc_srq_buf(dev
, pd
, srq
);
257 goto err_out_mailbox
;
259 spin_lock_init(&srq
->lock
);
261 init_waitqueue_head(&srq
->wait
);
262 mutex_init(&srq
->mutex
);
264 if (mthca_is_memfree(dev
))
265 mthca_arbel_init_srq_context(dev
, pd
, srq
, mailbox
->buf
);
267 mthca_tavor_init_srq_context(dev
, pd
, srq
, mailbox
->buf
);
269 err
= mthca_SW2HW_SRQ(dev
, mailbox
, srq
->srqn
, &status
);
272 mthca_warn(dev
, "SW2HW_SRQ failed (%d)\n", err
);
273 goto err_out_free_buf
;
276 mthca_warn(dev
, "SW2HW_SRQ returned status 0x%02x\n",
279 goto err_out_free_buf
;
282 spin_lock_irq(&dev
->srq_table
.lock
);
283 if (mthca_array_set(&dev
->srq_table
.srq
,
284 srq
->srqn
& (dev
->limits
.num_srqs
- 1),
286 spin_unlock_irq(&dev
->srq_table
.lock
);
287 goto err_out_free_srq
;
289 spin_unlock_irq(&dev
->srq_table
.lock
);
291 mthca_free_mailbox(dev
, mailbox
);
294 srq
->last_free
= srq
->max
- 1;
296 attr
->max_wr
= srq
->max
- 1;
297 attr
->max_sge
= srq
->max_gs
;
302 err
= mthca_HW2SW_SRQ(dev
, mailbox
, srq
->srqn
, &status
);
304 mthca_warn(dev
, "HW2SW_SRQ failed (%d)\n", err
);
306 mthca_warn(dev
, "HW2SW_SRQ returned status 0x%02x\n", status
);
309 if (!pd
->ibpd
.uobject
)
310 mthca_free_srq_buf(dev
, srq
);
313 mthca_free_mailbox(dev
, mailbox
);
316 if (!pd
->ibpd
.uobject
&& mthca_is_memfree(dev
))
317 mthca_free_db(dev
, MTHCA_DB_TYPE_SRQ
, srq
->db_index
);
320 mthca_table_put(dev
, dev
->srq_table
.table
, srq
->srqn
);
323 mthca_free(&dev
->srq_table
.alloc
, srq
->srqn
);
328 static inline int get_srq_refcount(struct mthca_dev
*dev
, struct mthca_srq
*srq
)
332 spin_lock_irq(&dev
->srq_table
.lock
);
334 spin_unlock_irq(&dev
->srq_table
.lock
);
339 void mthca_free_srq(struct mthca_dev
*dev
, struct mthca_srq
*srq
)
341 struct mthca_mailbox
*mailbox
;
345 mailbox
= mthca_alloc_mailbox(dev
, GFP_KERNEL
);
346 if (IS_ERR(mailbox
)) {
347 mthca_warn(dev
, "No memory for mailbox to free SRQ.\n");
351 err
= mthca_HW2SW_SRQ(dev
, mailbox
, srq
->srqn
, &status
);
353 mthca_warn(dev
, "HW2SW_SRQ failed (%d)\n", err
);
355 mthca_warn(dev
, "HW2SW_SRQ returned status 0x%02x\n", status
);
357 spin_lock_irq(&dev
->srq_table
.lock
);
358 mthca_array_clear(&dev
->srq_table
.srq
,
359 srq
->srqn
& (dev
->limits
.num_srqs
- 1));
361 spin_unlock_irq(&dev
->srq_table
.lock
);
363 wait_event(srq
->wait
, !get_srq_refcount(dev
, srq
));
365 if (!srq
->ibsrq
.uobject
) {
366 mthca_free_srq_buf(dev
, srq
);
367 if (mthca_is_memfree(dev
))
368 mthca_free_db(dev
, MTHCA_DB_TYPE_SRQ
, srq
->db_index
);
371 mthca_table_put(dev
, dev
->srq_table
.table
, srq
->srqn
);
372 mthca_free(&dev
->srq_table
.alloc
, srq
->srqn
);
373 mthca_free_mailbox(dev
, mailbox
);
376 int mthca_modify_srq(struct ib_srq
*ibsrq
, struct ib_srq_attr
*attr
,
377 enum ib_srq_attr_mask attr_mask
, struct ib_udata
*udata
)
379 struct mthca_dev
*dev
= to_mdev(ibsrq
->device
);
380 struct mthca_srq
*srq
= to_msrq(ibsrq
);
384 /* We don't support resizing SRQs (yet?) */
385 if (attr_mask
& IB_SRQ_MAX_WR
)
388 if (attr_mask
& IB_SRQ_LIMIT
) {
389 u32 max_wr
= mthca_is_memfree(dev
) ? srq
->max
- 1 : srq
->max
;
390 if (attr
->srq_limit
> max_wr
)
393 mutex_lock(&srq
->mutex
);
394 ret
= mthca_ARM_SRQ(dev
, srq
->srqn
, attr
->srq_limit
, &status
);
395 mutex_unlock(&srq
->mutex
);
406 int mthca_query_srq(struct ib_srq
*ibsrq
, struct ib_srq_attr
*srq_attr
)
408 struct mthca_dev
*dev
= to_mdev(ibsrq
->device
);
409 struct mthca_srq
*srq
= to_msrq(ibsrq
);
410 struct mthca_mailbox
*mailbox
;
411 struct mthca_arbel_srq_context
*arbel_ctx
;
412 struct mthca_tavor_srq_context
*tavor_ctx
;
416 mailbox
= mthca_alloc_mailbox(dev
, GFP_KERNEL
);
418 return PTR_ERR(mailbox
);
420 err
= mthca_QUERY_SRQ(dev
, srq
->srqn
, mailbox
, &status
);
424 if (mthca_is_memfree(dev
)) {
425 arbel_ctx
= mailbox
->buf
;
426 srq_attr
->srq_limit
= be16_to_cpu(arbel_ctx
->limit_watermark
);
428 tavor_ctx
= mailbox
->buf
;
429 srq_attr
->srq_limit
= be16_to_cpu(tavor_ctx
->limit_watermark
);
432 srq_attr
->max_wr
= srq
->max
- 1;
433 srq_attr
->max_sge
= srq
->max_gs
;
436 mthca_free_mailbox(dev
, mailbox
);
441 void mthca_srq_event(struct mthca_dev
*dev
, u32 srqn
,
442 enum ib_event_type event_type
)
444 struct mthca_srq
*srq
;
445 struct ib_event event
;
447 spin_lock(&dev
->srq_table
.lock
);
448 srq
= mthca_array_get(&dev
->srq_table
.srq
, srqn
& (dev
->limits
.num_srqs
- 1));
451 spin_unlock(&dev
->srq_table
.lock
);
454 mthca_warn(dev
, "Async event for bogus SRQ %08x\n", srqn
);
458 if (!srq
->ibsrq
.event_handler
)
461 event
.device
= &dev
->ib_dev
;
462 event
.event
= event_type
;
463 event
.element
.srq
= &srq
->ibsrq
;
464 srq
->ibsrq
.event_handler(&event
, srq
->ibsrq
.srq_context
);
467 spin_lock(&dev
->srq_table
.lock
);
468 if (!--srq
->refcount
)
470 spin_unlock(&dev
->srq_table
.lock
);
474 * This function must be called with IRQs disabled.
476 void mthca_free_srq_wqe(struct mthca_srq
*srq
, u32 wqe_addr
)
479 struct mthca_next_seg
*last_free
;
481 ind
= wqe_addr
>> srq
->wqe_shift
;
483 spin_lock(&srq
->lock
);
485 last_free
= get_wqe(srq
, srq
->last_free
);
486 *wqe_to_link(last_free
) = ind
;
487 last_free
->nda_op
= htonl((ind
<< srq
->wqe_shift
) | 1);
488 *wqe_to_link(get_wqe(srq
, ind
)) = -1;
489 srq
->last_free
= ind
;
491 spin_unlock(&srq
->lock
);
494 int mthca_tavor_post_srq_recv(struct ib_srq
*ibsrq
, struct ib_recv_wr
*wr
,
495 struct ib_recv_wr
**bad_wr
)
497 struct mthca_dev
*dev
= to_mdev(ibsrq
->device
);
498 struct mthca_srq
*srq
= to_msrq(ibsrq
);
509 spin_lock_irqsave(&srq
->lock
, flags
);
511 first_ind
= srq
->first_free
;
513 for (nreq
= 0; wr
; wr
= wr
->next
) {
514 ind
= srq
->first_free
;
515 wqe
= get_wqe(srq
, ind
);
516 next_ind
= *wqe_to_link(wqe
);
518 if (unlikely(next_ind
< 0)) {
519 mthca_err(dev
, "SRQ %06x full\n", srq
->srqn
);
525 prev_wqe
= srq
->last
;
528 ((struct mthca_next_seg
*) wqe
)->ee_nds
= 0;
529 /* flags field will always remain 0 */
531 wqe
+= sizeof (struct mthca_next_seg
);
533 if (unlikely(wr
->num_sge
> srq
->max_gs
)) {
536 srq
->last
= prev_wqe
;
540 for (i
= 0; i
< wr
->num_sge
; ++i
) {
541 mthca_set_data_seg(wqe
, wr
->sg_list
+ i
);
542 wqe
+= sizeof (struct mthca_data_seg
);
546 mthca_set_data_seg_inval(wqe
);
548 ((struct mthca_next_seg
*) prev_wqe
)->ee_nds
=
549 cpu_to_be32(MTHCA_NEXT_DBD
);
551 srq
->wrid
[ind
] = wr
->wr_id
;
552 srq
->first_free
= next_ind
;
555 if (unlikely(nreq
== MTHCA_TAVOR_MAX_WQES_PER_RECV_DB
)) {
559 * Make sure that descriptors are written
560 * before doorbell is rung.
564 mthca_write64(first_ind
<< srq
->wqe_shift
, srq
->srqn
<< 8,
565 dev
->kar
+ MTHCA_RECEIVE_DOORBELL
,
566 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
568 first_ind
= srq
->first_free
;
574 * Make sure that descriptors are written before
579 mthca_write64(first_ind
<< srq
->wqe_shift
, (srq
->srqn
<< 8) | nreq
,
580 dev
->kar
+ MTHCA_RECEIVE_DOORBELL
,
581 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
585 * Make sure doorbells don't leak out of SRQ spinlock and
586 * reach the HCA out of order:
590 spin_unlock_irqrestore(&srq
->lock
, flags
);
594 int mthca_arbel_post_srq_recv(struct ib_srq
*ibsrq
, struct ib_recv_wr
*wr
,
595 struct ib_recv_wr
**bad_wr
)
597 struct mthca_dev
*dev
= to_mdev(ibsrq
->device
);
598 struct mthca_srq
*srq
= to_msrq(ibsrq
);
607 spin_lock_irqsave(&srq
->lock
, flags
);
609 for (nreq
= 0; wr
; ++nreq
, wr
= wr
->next
) {
610 ind
= srq
->first_free
;
611 wqe
= get_wqe(srq
, ind
);
612 next_ind
= *wqe_to_link(wqe
);
614 if (unlikely(next_ind
< 0)) {
615 mthca_err(dev
, "SRQ %06x full\n", srq
->srqn
);
621 ((struct mthca_next_seg
*) wqe
)->ee_nds
= 0;
622 /* flags field will always remain 0 */
624 wqe
+= sizeof (struct mthca_next_seg
);
626 if (unlikely(wr
->num_sge
> srq
->max_gs
)) {
632 for (i
= 0; i
< wr
->num_sge
; ++i
) {
633 mthca_set_data_seg(wqe
, wr
->sg_list
+ i
);
634 wqe
+= sizeof (struct mthca_data_seg
);
638 mthca_set_data_seg_inval(wqe
);
640 srq
->wrid
[ind
] = wr
->wr_id
;
641 srq
->first_free
= next_ind
;
645 srq
->counter
+= nreq
;
648 * Make sure that descriptors are written before
649 * we write doorbell record.
652 *srq
->db
= cpu_to_be32(srq
->counter
);
655 spin_unlock_irqrestore(&srq
->lock
, flags
);
659 int mthca_max_srq_sge(struct mthca_dev
*dev
)
661 if (mthca_is_memfree(dev
))
662 return dev
->limits
.max_sg
;
665 * SRQ allocations are based on powers of 2 for Tavor,
666 * (although they only need to be multiples of 16 bytes).
668 * Therefore, we need to base the max number of sg entries on
669 * the largest power of 2 descriptor size that is <= to the
670 * actual max WQE descriptor size, rather than return the
671 * max_sg value given by the firmware (which is based on WQE
672 * sizes as multiples of 16, not powers of 2).
674 * If SRQ implementation is changed for Tavor to be based on
675 * multiples of 16, the calculation below can be deleted and
676 * the FW max_sg value returned.
678 return min_t(int, dev
->limits
.max_sg
,
679 ((1 << (fls(dev
->limits
.max_desc_sz
) - 1)) -
680 sizeof (struct mthca_next_seg
)) /
681 sizeof (struct mthca_data_seg
));
684 int mthca_init_srq_table(struct mthca_dev
*dev
)
688 if (!(dev
->mthca_flags
& MTHCA_FLAG_SRQ
))
691 spin_lock_init(&dev
->srq_table
.lock
);
693 err
= mthca_alloc_init(&dev
->srq_table
.alloc
,
694 dev
->limits
.num_srqs
,
695 dev
->limits
.num_srqs
- 1,
696 dev
->limits
.reserved_srqs
);
700 err
= mthca_array_init(&dev
->srq_table
.srq
,
701 dev
->limits
.num_srqs
);
703 mthca_alloc_cleanup(&dev
->srq_table
.alloc
);
708 void mthca_cleanup_srq_table(struct mthca_dev
*dev
)
710 if (!(dev
->mthca_flags
& MTHCA_FLAG_SRQ
))
713 mthca_array_cleanup(&dev
->srq_table
.srq
, dev
->limits
.num_srqs
);
714 mthca_alloc_cleanup(&dev
->srq_table
.alloc
);