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>
40 #include "mthca_dev.h"
41 #include "mthca_cmd.h"
42 #include "mthca_memfree.h"
43 #include "mthca_wqe.h"
46 MTHCA_MAX_DIRECT_SRQ_SIZE
= 4 * PAGE_SIZE
49 struct mthca_tavor_srq_context
{
50 __be64 wqe_base_ds
; /* low 6 bits is descriptor size */
54 __be16 limit_watermark
;
59 struct mthca_arbel_srq_context
{
60 __be32 state_logsize_srqn
;
63 __be32 logstride_usrpage
;
66 __be16 limit_watermark
;
73 static void *get_wqe(struct mthca_srq
*srq
, int n
)
76 return srq
->queue
.direct
.buf
+ (n
<< srq
->wqe_shift
);
78 return srq
->queue
.page_list
[(n
<< srq
->wqe_shift
) >> PAGE_SHIFT
].buf
+
79 ((n
<< srq
->wqe_shift
) & (PAGE_SIZE
- 1));
83 * Return a pointer to the location within a WQE that we're using as a
84 * link when the WQE is in the free list. We use the imm field
85 * because in the Tavor case, posting a WQE may overwrite the next
86 * segment of the previous WQE, but a receive WQE will never touch the
87 * imm field. This avoids corrupting our free list if the previous
88 * WQE has already completed and been put on the free list when we
91 static inline int *wqe_to_link(void *wqe
)
93 return (int *) (wqe
+ offsetof(struct mthca_next_seg
, imm
));
96 static void mthca_tavor_init_srq_context(struct mthca_dev
*dev
,
98 struct mthca_srq
*srq
,
99 struct mthca_tavor_srq_context
*context
)
101 memset(context
, 0, sizeof *context
);
103 context
->wqe_base_ds
= cpu_to_be64(1 << (srq
->wqe_shift
- 4));
104 context
->state_pd
= cpu_to_be32(pd
->pd_num
);
105 context
->lkey
= cpu_to_be32(srq
->mr
.ibmr
.lkey
);
107 if (pd
->ibpd
.uobject
)
109 cpu_to_be32(to_mucontext(pd
->ibpd
.uobject
->context
)->uar
.index
);
111 context
->uar
= cpu_to_be32(dev
->driver_uar
.index
);
114 static void mthca_arbel_init_srq_context(struct mthca_dev
*dev
,
116 struct mthca_srq
*srq
,
117 struct mthca_arbel_srq_context
*context
)
121 memset(context
, 0, sizeof *context
);
124 * Put max in a temporary variable to work around gcc bug
125 * triggered by ilog2() on sparc64.
128 logsize
= ilog2(max
);
129 context
->state_logsize_srqn
= cpu_to_be32(logsize
<< 24 | srq
->srqn
);
130 context
->lkey
= cpu_to_be32(srq
->mr
.ibmr
.lkey
);
131 context
->db_index
= cpu_to_be32(srq
->db_index
);
132 context
->logstride_usrpage
= cpu_to_be32((srq
->wqe_shift
- 4) << 29);
133 if (pd
->ibpd
.uobject
)
134 context
->logstride_usrpage
|=
135 cpu_to_be32(to_mucontext(pd
->ibpd
.uobject
->context
)->uar
.index
);
137 context
->logstride_usrpage
|= cpu_to_be32(dev
->driver_uar
.index
);
138 context
->eq_pd
= cpu_to_be32(MTHCA_EQ_ASYNC
<< 24 | pd
->pd_num
);
141 static void mthca_free_srq_buf(struct mthca_dev
*dev
, struct mthca_srq
*srq
)
143 mthca_buf_free(dev
, srq
->max
<< srq
->wqe_shift
, &srq
->queue
,
144 srq
->is_direct
, &srq
->mr
);
148 static int mthca_alloc_srq_buf(struct mthca_dev
*dev
, struct mthca_pd
*pd
,
149 struct mthca_srq
*srq
)
151 struct mthca_data_seg
*scatter
;
156 if (pd
->ibpd
.uobject
)
159 srq
->wrid
= kmalloc(srq
->max
* sizeof (u64
), GFP_KERNEL
);
163 err
= mthca_buf_alloc(dev
, srq
->max
<< srq
->wqe_shift
,
164 MTHCA_MAX_DIRECT_SRQ_SIZE
,
165 &srq
->queue
, &srq
->is_direct
, pd
, 1, &srq
->mr
);
172 * Now initialize the SRQ buffer so that all of the WQEs are
173 * linked into the list of free WQEs. In addition, set the
174 * scatter list L_Keys to the sentry value of 0x100.
176 for (i
= 0; i
< srq
->max
; ++i
) {
177 wqe
= get_wqe(srq
, i
);
179 *wqe_to_link(wqe
) = i
< srq
->max
- 1 ? i
+ 1 : -1;
181 for (scatter
= wqe
+ sizeof (struct mthca_next_seg
);
182 (void *) scatter
< wqe
+ (1 << srq
->wqe_shift
);
184 scatter
->lkey
= cpu_to_be32(MTHCA_INVAL_LKEY
);
187 srq
->last
= get_wqe(srq
, srq
->max
- 1);
192 int mthca_alloc_srq(struct mthca_dev
*dev
, struct mthca_pd
*pd
,
193 struct ib_srq_attr
*attr
, struct mthca_srq
*srq
)
195 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
, &status
);
265 mthca_warn(dev
, "SW2HW_SRQ failed (%d)\n", err
);
266 goto err_out_free_buf
;
269 mthca_warn(dev
, "SW2HW_SRQ returned status 0x%02x\n",
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
, &status
);
297 mthca_warn(dev
, "HW2SW_SRQ failed (%d)\n", err
);
299 mthca_warn(dev
, "HW2SW_SRQ returned status 0x%02x\n", status
);
302 if (!pd
->ibpd
.uobject
)
303 mthca_free_srq_buf(dev
, srq
);
306 mthca_free_mailbox(dev
, mailbox
);
309 if (!pd
->ibpd
.uobject
&& mthca_is_memfree(dev
))
310 mthca_free_db(dev
, MTHCA_DB_TYPE_SRQ
, srq
->db_index
);
313 mthca_table_put(dev
, dev
->srq_table
.table
, srq
->srqn
);
316 mthca_free(&dev
->srq_table
.alloc
, srq
->srqn
);
321 static inline int get_srq_refcount(struct mthca_dev
*dev
, struct mthca_srq
*srq
)
325 spin_lock_irq(&dev
->srq_table
.lock
);
327 spin_unlock_irq(&dev
->srq_table
.lock
);
332 void mthca_free_srq(struct mthca_dev
*dev
, struct mthca_srq
*srq
)
334 struct mthca_mailbox
*mailbox
;
338 mailbox
= mthca_alloc_mailbox(dev
, GFP_KERNEL
);
339 if (IS_ERR(mailbox
)) {
340 mthca_warn(dev
, "No memory for mailbox to free SRQ.\n");
344 err
= mthca_HW2SW_SRQ(dev
, mailbox
, srq
->srqn
, &status
);
346 mthca_warn(dev
, "HW2SW_SRQ failed (%d)\n", err
);
348 mthca_warn(dev
, "HW2SW_SRQ returned status 0x%02x\n", status
);
350 spin_lock_irq(&dev
->srq_table
.lock
);
351 mthca_array_clear(&dev
->srq_table
.srq
,
352 srq
->srqn
& (dev
->limits
.num_srqs
- 1));
354 spin_unlock_irq(&dev
->srq_table
.lock
);
356 wait_event(srq
->wait
, !get_srq_refcount(dev
, srq
));
358 if (!srq
->ibsrq
.uobject
) {
359 mthca_free_srq_buf(dev
, srq
);
360 if (mthca_is_memfree(dev
))
361 mthca_free_db(dev
, MTHCA_DB_TYPE_SRQ
, srq
->db_index
);
364 mthca_table_put(dev
, dev
->srq_table
.table
, srq
->srqn
);
365 mthca_free(&dev
->srq_table
.alloc
, srq
->srqn
);
366 mthca_free_mailbox(dev
, mailbox
);
369 int mthca_modify_srq(struct ib_srq
*ibsrq
, struct ib_srq_attr
*attr
,
370 enum ib_srq_attr_mask attr_mask
, struct ib_udata
*udata
)
372 struct mthca_dev
*dev
= to_mdev(ibsrq
->device
);
373 struct mthca_srq
*srq
= to_msrq(ibsrq
);
377 /* We don't support resizing SRQs (yet?) */
378 if (attr_mask
& IB_SRQ_MAX_WR
)
381 if (attr_mask
& IB_SRQ_LIMIT
) {
382 u32 max_wr
= mthca_is_memfree(dev
) ? srq
->max
- 1 : srq
->max
;
383 if (attr
->srq_limit
> max_wr
)
386 mutex_lock(&srq
->mutex
);
387 ret
= mthca_ARM_SRQ(dev
, srq
->srqn
, attr
->srq_limit
, &status
);
388 mutex_unlock(&srq
->mutex
);
399 int mthca_query_srq(struct ib_srq
*ibsrq
, struct ib_srq_attr
*srq_attr
)
401 struct mthca_dev
*dev
= to_mdev(ibsrq
->device
);
402 struct mthca_srq
*srq
= to_msrq(ibsrq
);
403 struct mthca_mailbox
*mailbox
;
404 struct mthca_arbel_srq_context
*arbel_ctx
;
405 struct mthca_tavor_srq_context
*tavor_ctx
;
409 mailbox
= mthca_alloc_mailbox(dev
, GFP_KERNEL
);
411 return PTR_ERR(mailbox
);
413 err
= mthca_QUERY_SRQ(dev
, srq
->srqn
, mailbox
, &status
);
417 if (mthca_is_memfree(dev
)) {
418 arbel_ctx
= mailbox
->buf
;
419 srq_attr
->srq_limit
= be16_to_cpu(arbel_ctx
->limit_watermark
);
421 tavor_ctx
= mailbox
->buf
;
422 srq_attr
->srq_limit
= be16_to_cpu(tavor_ctx
->limit_watermark
);
425 srq_attr
->max_wr
= srq
->max
- 1;
426 srq_attr
->max_sge
= srq
->max_gs
;
429 mthca_free_mailbox(dev
, mailbox
);
434 void mthca_srq_event(struct mthca_dev
*dev
, u32 srqn
,
435 enum ib_event_type event_type
)
437 struct mthca_srq
*srq
;
438 struct ib_event event
;
440 spin_lock(&dev
->srq_table
.lock
);
441 srq
= mthca_array_get(&dev
->srq_table
.srq
, srqn
& (dev
->limits
.num_srqs
- 1));
444 spin_unlock(&dev
->srq_table
.lock
);
447 mthca_warn(dev
, "Async event for bogus SRQ %08x\n", srqn
);
451 if (!srq
->ibsrq
.event_handler
)
454 event
.device
= &dev
->ib_dev
;
455 event
.event
= event_type
;
456 event
.element
.srq
= &srq
->ibsrq
;
457 srq
->ibsrq
.event_handler(&event
, srq
->ibsrq
.srq_context
);
460 spin_lock(&dev
->srq_table
.lock
);
461 if (!--srq
->refcount
)
463 spin_unlock(&dev
->srq_table
.lock
);
467 * This function must be called with IRQs disabled.
469 void mthca_free_srq_wqe(struct mthca_srq
*srq
, u32 wqe_addr
)
473 ind
= wqe_addr
>> srq
->wqe_shift
;
475 spin_lock(&srq
->lock
);
477 if (likely(srq
->first_free
>= 0))
478 *wqe_to_link(get_wqe(srq
, srq
->last_free
)) = ind
;
480 srq
->first_free
= ind
;
482 *wqe_to_link(get_wqe(srq
, ind
)) = -1;
483 srq
->last_free
= ind
;
485 spin_unlock(&srq
->lock
);
488 int mthca_tavor_post_srq_recv(struct ib_srq
*ibsrq
, struct ib_recv_wr
*wr
,
489 struct ib_recv_wr
**bad_wr
)
491 struct mthca_dev
*dev
= to_mdev(ibsrq
->device
);
492 struct mthca_srq
*srq
= to_msrq(ibsrq
);
504 spin_lock_irqsave(&srq
->lock
, flags
);
506 first_ind
= srq
->first_free
;
508 for (nreq
= 0; wr
; wr
= wr
->next
) {
509 ind
= srq
->first_free
;
512 mthca_err(dev
, "SRQ %06x full\n", srq
->srqn
);
518 wqe
= get_wqe(srq
, ind
);
519 next_ind
= *wqe_to_link(wqe
);
522 mthca_err(dev
, "SRQ %06x full\n", srq
->srqn
);
528 prev_wqe
= srq
->last
;
531 ((struct mthca_next_seg
*) wqe
)->nda_op
= 0;
532 ((struct mthca_next_seg
*) wqe
)->ee_nds
= 0;
533 /* flags field will always remain 0 */
535 wqe
+= sizeof (struct mthca_next_seg
);
537 if (unlikely(wr
->num_sge
> srq
->max_gs
)) {
540 srq
->last
= prev_wqe
;
544 for (i
= 0; i
< wr
->num_sge
; ++i
) {
545 ((struct mthca_data_seg
*) wqe
)->byte_count
=
546 cpu_to_be32(wr
->sg_list
[i
].length
);
547 ((struct mthca_data_seg
*) wqe
)->lkey
=
548 cpu_to_be32(wr
->sg_list
[i
].lkey
);
549 ((struct mthca_data_seg
*) wqe
)->addr
=
550 cpu_to_be64(wr
->sg_list
[i
].addr
);
551 wqe
+= sizeof (struct mthca_data_seg
);
554 if (i
< srq
->max_gs
) {
555 ((struct mthca_data_seg
*) wqe
)->byte_count
= 0;
556 ((struct mthca_data_seg
*) wqe
)->lkey
= cpu_to_be32(MTHCA_INVAL_LKEY
);
557 ((struct mthca_data_seg
*) wqe
)->addr
= 0;
560 ((struct mthca_next_seg
*) prev_wqe
)->nda_op
=
561 cpu_to_be32((ind
<< srq
->wqe_shift
) | 1);
563 ((struct mthca_next_seg
*) prev_wqe
)->ee_nds
=
564 cpu_to_be32(MTHCA_NEXT_DBD
);
566 srq
->wrid
[ind
] = wr
->wr_id
;
567 srq
->first_free
= next_ind
;
570 if (unlikely(nreq
== MTHCA_TAVOR_MAX_WQES_PER_RECV_DB
)) {
573 doorbell
[0] = cpu_to_be32(first_ind
<< srq
->wqe_shift
);
574 doorbell
[1] = cpu_to_be32(srq
->srqn
<< 8);
577 * Make sure that descriptors are written
578 * before doorbell is rung.
582 mthca_write64(doorbell
,
583 dev
->kar
+ MTHCA_RECEIVE_DOORBELL
,
584 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
586 first_ind
= srq
->first_free
;
591 doorbell
[0] = cpu_to_be32(first_ind
<< srq
->wqe_shift
);
592 doorbell
[1] = cpu_to_be32((srq
->srqn
<< 8) | nreq
);
595 * Make sure that descriptors are written before
600 mthca_write64(doorbell
,
601 dev
->kar
+ MTHCA_RECEIVE_DOORBELL
,
602 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
606 * Make sure doorbells don't leak out of SRQ spinlock and
607 * reach the HCA out of order:
611 spin_unlock_irqrestore(&srq
->lock
, flags
);
615 int mthca_arbel_post_srq_recv(struct ib_srq
*ibsrq
, struct ib_recv_wr
*wr
,
616 struct ib_recv_wr
**bad_wr
)
618 struct mthca_dev
*dev
= to_mdev(ibsrq
->device
);
619 struct mthca_srq
*srq
= to_msrq(ibsrq
);
628 spin_lock_irqsave(&srq
->lock
, flags
);
630 for (nreq
= 0; wr
; ++nreq
, wr
= wr
->next
) {
631 ind
= srq
->first_free
;
634 mthca_err(dev
, "SRQ %06x full\n", srq
->srqn
);
640 wqe
= get_wqe(srq
, ind
);
641 next_ind
= *wqe_to_link(wqe
);
644 mthca_err(dev
, "SRQ %06x full\n", srq
->srqn
);
650 ((struct mthca_next_seg
*) wqe
)->nda_op
=
651 cpu_to_be32((next_ind
<< srq
->wqe_shift
) | 1);
652 ((struct mthca_next_seg
*) wqe
)->ee_nds
= 0;
653 /* flags field will always remain 0 */
655 wqe
+= sizeof (struct mthca_next_seg
);
657 if (unlikely(wr
->num_sge
> srq
->max_gs
)) {
663 for (i
= 0; i
< wr
->num_sge
; ++i
) {
664 ((struct mthca_data_seg
*) wqe
)->byte_count
=
665 cpu_to_be32(wr
->sg_list
[i
].length
);
666 ((struct mthca_data_seg
*) wqe
)->lkey
=
667 cpu_to_be32(wr
->sg_list
[i
].lkey
);
668 ((struct mthca_data_seg
*) wqe
)->addr
=
669 cpu_to_be64(wr
->sg_list
[i
].addr
);
670 wqe
+= sizeof (struct mthca_data_seg
);
673 if (i
< srq
->max_gs
) {
674 ((struct mthca_data_seg
*) wqe
)->byte_count
= 0;
675 ((struct mthca_data_seg
*) wqe
)->lkey
= cpu_to_be32(MTHCA_INVAL_LKEY
);
676 ((struct mthca_data_seg
*) wqe
)->addr
= 0;
679 srq
->wrid
[ind
] = wr
->wr_id
;
680 srq
->first_free
= next_ind
;
684 srq
->counter
+= nreq
;
687 * Make sure that descriptors are written before
688 * we write doorbell record.
691 *srq
->db
= cpu_to_be32(srq
->counter
);
694 spin_unlock_irqrestore(&srq
->lock
, flags
);
698 int mthca_max_srq_sge(struct mthca_dev
*dev
)
700 if (mthca_is_memfree(dev
))
701 return dev
->limits
.max_sg
;
704 * SRQ allocations are based on powers of 2 for Tavor,
705 * (although they only need to be multiples of 16 bytes).
707 * Therefore, we need to base the max number of sg entries on
708 * the largest power of 2 descriptor size that is <= to the
709 * actual max WQE descriptor size, rather than return the
710 * max_sg value given by the firmware (which is based on WQE
711 * sizes as multiples of 16, not powers of 2).
713 * If SRQ implementation is changed for Tavor to be based on
714 * multiples of 16, the calculation below can be deleted and
715 * the FW max_sg value returned.
717 return min_t(int, dev
->limits
.max_sg
,
718 ((1 << (fls(dev
->limits
.max_desc_sz
) - 1)) -
719 sizeof (struct mthca_next_seg
)) /
720 sizeof (struct mthca_data_seg
));
723 int mthca_init_srq_table(struct mthca_dev
*dev
)
727 if (!(dev
->mthca_flags
& MTHCA_FLAG_SRQ
))
730 spin_lock_init(&dev
->srq_table
.lock
);
732 err
= mthca_alloc_init(&dev
->srq_table
.alloc
,
733 dev
->limits
.num_srqs
,
734 dev
->limits
.num_srqs
- 1,
735 dev
->limits
.reserved_srqs
);
739 err
= mthca_array_init(&dev
->srq_table
.srq
,
740 dev
->limits
.num_srqs
);
742 mthca_alloc_cleanup(&dev
->srq_table
.alloc
);
747 void mthca_cleanup_srq_table(struct mthca_dev
*dev
)
749 if (!(dev
->mthca_flags
& MTHCA_FLAG_SRQ
))
752 mthca_array_cleanup(&dev
->srq_table
.srq
, dev
->limits
.num_srqs
);
753 mthca_alloc_cleanup(&dev
->srq_table
.alloc
);