2 * Copyright (c) 2006 Chelsio, Inc. 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 #include <linux/sched.h>
33 #include <linux/gfp.h>
34 #include "iwch_provider.h"
38 #include "cxio_resource.h"
42 static int build_rdma_send(union t3_wr
*wqe
, struct ib_send_wr
*wr
,
50 if (wr
->send_flags
& IB_SEND_SOLICITED
)
51 wqe
->send
.rdmaop
= T3_SEND_WITH_SE
;
53 wqe
->send
.rdmaop
= T3_SEND
;
54 wqe
->send
.rem_stag
= 0;
56 case IB_WR_SEND_WITH_INV
:
57 if (wr
->send_flags
& IB_SEND_SOLICITED
)
58 wqe
->send
.rdmaop
= T3_SEND_WITH_SE_INV
;
60 wqe
->send
.rdmaop
= T3_SEND_WITH_INV
;
61 wqe
->send
.rem_stag
= cpu_to_be32(wr
->ex
.invalidate_rkey
);
66 if (wr
->num_sge
> T3_MAX_SGE
)
68 wqe
->send
.reserved
[0] = 0;
69 wqe
->send
.reserved
[1] = 0;
70 wqe
->send
.reserved
[2] = 0;
72 for (i
= 0; i
< wr
->num_sge
; i
++) {
73 if ((plen
+ wr
->sg_list
[i
].length
) < plen
)
76 plen
+= wr
->sg_list
[i
].length
;
77 wqe
->send
.sgl
[i
].stag
= cpu_to_be32(wr
->sg_list
[i
].lkey
);
78 wqe
->send
.sgl
[i
].len
= cpu_to_be32(wr
->sg_list
[i
].length
);
79 wqe
->send
.sgl
[i
].to
= cpu_to_be64(wr
->sg_list
[i
].addr
);
81 wqe
->send
.num_sgle
= cpu_to_be32(wr
->num_sge
);
82 *flit_cnt
= 4 + ((wr
->num_sge
) << 1);
83 wqe
->send
.plen
= cpu_to_be32(plen
);
87 static int build_rdma_write(union t3_wr
*wqe
, struct ib_send_wr
*wr
,
92 if (wr
->num_sge
> T3_MAX_SGE
)
94 wqe
->write
.rdmaop
= T3_RDMA_WRITE
;
95 wqe
->write
.reserved
[0] = 0;
96 wqe
->write
.reserved
[1] = 0;
97 wqe
->write
.reserved
[2] = 0;
98 wqe
->write
.stag_sink
= cpu_to_be32(wr
->wr
.rdma
.rkey
);
99 wqe
->write
.to_sink
= cpu_to_be64(wr
->wr
.rdma
.remote_addr
);
101 if (wr
->opcode
== IB_WR_RDMA_WRITE_WITH_IMM
) {
103 wqe
->write
.sgl
[0].stag
= wr
->ex
.imm_data
;
104 wqe
->write
.sgl
[0].len
= cpu_to_be32(0);
105 wqe
->write
.num_sgle
= cpu_to_be32(0);
109 for (i
= 0; i
< wr
->num_sge
; i
++) {
110 if ((plen
+ wr
->sg_list
[i
].length
) < plen
) {
113 plen
+= wr
->sg_list
[i
].length
;
114 wqe
->write
.sgl
[i
].stag
=
115 cpu_to_be32(wr
->sg_list
[i
].lkey
);
116 wqe
->write
.sgl
[i
].len
=
117 cpu_to_be32(wr
->sg_list
[i
].length
);
118 wqe
->write
.sgl
[i
].to
=
119 cpu_to_be64(wr
->sg_list
[i
].addr
);
121 wqe
->write
.num_sgle
= cpu_to_be32(wr
->num_sge
);
122 *flit_cnt
= 5 + ((wr
->num_sge
) << 1);
124 wqe
->write
.plen
= cpu_to_be32(plen
);
128 static int build_rdma_read(union t3_wr
*wqe
, struct ib_send_wr
*wr
,
133 wqe
->read
.rdmaop
= T3_READ_REQ
;
134 if (wr
->opcode
== IB_WR_RDMA_READ_WITH_INV
)
135 wqe
->read
.local_inv
= 1;
137 wqe
->read
.local_inv
= 0;
138 wqe
->read
.reserved
[0] = 0;
139 wqe
->read
.reserved
[1] = 0;
140 wqe
->read
.rem_stag
= cpu_to_be32(wr
->wr
.rdma
.rkey
);
141 wqe
->read
.rem_to
= cpu_to_be64(wr
->wr
.rdma
.remote_addr
);
142 wqe
->read
.local_stag
= cpu_to_be32(wr
->sg_list
[0].lkey
);
143 wqe
->read
.local_len
= cpu_to_be32(wr
->sg_list
[0].length
);
144 wqe
->read
.local_to
= cpu_to_be64(wr
->sg_list
[0].addr
);
145 *flit_cnt
= sizeof(struct t3_rdma_read_wr
) >> 3;
149 static int build_fastreg(union t3_wr
*wqe
, struct ib_send_wr
*wr
,
150 u8
*flit_cnt
, int *wr_cnt
, struct t3_wq
*wq
)
155 if (wr
->wr
.fast_reg
.page_list_len
> T3_MAX_FASTREG_DEPTH
)
158 wqe
->fastreg
.stag
= cpu_to_be32(wr
->wr
.fast_reg
.rkey
);
159 wqe
->fastreg
.len
= cpu_to_be32(wr
->wr
.fast_reg
.length
);
160 wqe
->fastreg
.va_base_hi
= cpu_to_be32(wr
->wr
.fast_reg
.iova_start
>> 32);
161 wqe
->fastreg
.va_base_lo_fbo
=
162 cpu_to_be32(wr
->wr
.fast_reg
.iova_start
& 0xffffffff);
163 wqe
->fastreg
.page_type_perms
= cpu_to_be32(
164 V_FR_PAGE_COUNT(wr
->wr
.fast_reg
.page_list_len
) |
165 V_FR_PAGE_SIZE(wr
->wr
.fast_reg
.page_shift
-12) |
166 V_FR_TYPE(TPT_VATO
) |
167 V_FR_PERMS(iwch_ib_to_tpt_access(wr
->wr
.fast_reg
.access_flags
)));
168 p
= &wqe
->fastreg
.pbl_addrs
[0];
169 for (i
= 0; i
< wr
->wr
.fast_reg
.page_list_len
; i
++, p
++) {
171 /* If we need a 2nd WR, then set it up */
172 if (i
== T3_MAX_FASTREG_FRAG
) {
174 wqe
= (union t3_wr
*)(wq
->queue
+
175 Q_PTR2IDX((wq
->wptr
+1), wq
->size_log2
));
176 build_fw_riwrh((void *)wqe
, T3_WR_FASTREG
, 0,
177 Q_GENBIT(wq
->wptr
+ 1, wq
->size_log2
),
178 0, 1 + wr
->wr
.fast_reg
.page_list_len
- T3_MAX_FASTREG_FRAG
,
181 p
= &wqe
->pbl_frag
.pbl_addrs
[0];
183 *p
= cpu_to_be64((u64
)wr
->wr
.fast_reg
.page_list
->page_list
[i
]);
185 *flit_cnt
= 5 + wr
->wr
.fast_reg
.page_list_len
;
191 static int build_inv_stag(union t3_wr
*wqe
, struct ib_send_wr
*wr
,
194 wqe
->local_inv
.stag
= cpu_to_be32(wr
->ex
.invalidate_rkey
);
195 wqe
->local_inv
.reserved
= 0;
196 *flit_cnt
= sizeof(struct t3_local_inv_wr
) >> 3;
200 static int iwch_sgl2pbl_map(struct iwch_dev
*rhp
, struct ib_sge
*sg_list
,
201 u32 num_sgle
, u32
* pbl_addr
, u8
* page_size
)
206 for (i
= 0; i
< num_sgle
; i
++) {
208 mhp
= get_mhp(rhp
, (sg_list
[i
].lkey
) >> 8);
210 PDBG("%s %d\n", __func__
, __LINE__
);
213 if (!mhp
->attr
.state
) {
214 PDBG("%s %d\n", __func__
, __LINE__
);
217 if (mhp
->attr
.zbva
) {
218 PDBG("%s %d\n", __func__
, __LINE__
);
222 if (sg_list
[i
].addr
< mhp
->attr
.va_fbo
) {
223 PDBG("%s %d\n", __func__
, __LINE__
);
226 if (sg_list
[i
].addr
+ ((u64
) sg_list
[i
].length
) <
228 PDBG("%s %d\n", __func__
, __LINE__
);
231 if (sg_list
[i
].addr
+ ((u64
) sg_list
[i
].length
) >
232 mhp
->attr
.va_fbo
+ ((u64
) mhp
->attr
.len
)) {
233 PDBG("%s %d\n", __func__
, __LINE__
);
236 offset
= sg_list
[i
].addr
- mhp
->attr
.va_fbo
;
237 offset
+= mhp
->attr
.va_fbo
&
238 ((1UL << (12 + mhp
->attr
.page_size
)) - 1);
239 pbl_addr
[i
] = ((mhp
->attr
.pbl_addr
-
240 rhp
->rdev
.rnic_info
.pbl_base
) >> 3) +
241 (offset
>> (12 + mhp
->attr
.page_size
));
242 page_size
[i
] = mhp
->attr
.page_size
;
247 static int build_rdma_recv(struct iwch_qp
*qhp
, union t3_wr
*wqe
,
248 struct ib_recv_wr
*wr
)
251 u32 pbl_addr
[T3_MAX_SGE
];
252 u8 page_size
[T3_MAX_SGE
];
254 err
= iwch_sgl2pbl_map(qhp
->rhp
, wr
->sg_list
, wr
->num_sge
, pbl_addr
,
258 wqe
->recv
.pagesz
[0] = page_size
[0];
259 wqe
->recv
.pagesz
[1] = page_size
[1];
260 wqe
->recv
.pagesz
[2] = page_size
[2];
261 wqe
->recv
.pagesz
[3] = page_size
[3];
262 wqe
->recv
.num_sgle
= cpu_to_be32(wr
->num_sge
);
263 for (i
= 0; i
< wr
->num_sge
; i
++) {
264 wqe
->recv
.sgl
[i
].stag
= cpu_to_be32(wr
->sg_list
[i
].lkey
);
265 wqe
->recv
.sgl
[i
].len
= cpu_to_be32(wr
->sg_list
[i
].length
);
267 /* to in the WQE == the offset into the page */
268 wqe
->recv
.sgl
[i
].to
= cpu_to_be64(((u32
)wr
->sg_list
[i
].addr
) &
269 ((1UL << (12 + page_size
[i
])) - 1));
271 /* pbl_addr is the adapters address in the PBL */
272 wqe
->recv
.pbl_addr
[i
] = cpu_to_be32(pbl_addr
[i
]);
274 for (; i
< T3_MAX_SGE
; i
++) {
275 wqe
->recv
.sgl
[i
].stag
= 0;
276 wqe
->recv
.sgl
[i
].len
= 0;
277 wqe
->recv
.sgl
[i
].to
= 0;
278 wqe
->recv
.pbl_addr
[i
] = 0;
280 qhp
->wq
.rq
[Q_PTR2IDX(qhp
->wq
.rq_wptr
,
281 qhp
->wq
.rq_size_log2
)].wr_id
= wr
->wr_id
;
282 qhp
->wq
.rq
[Q_PTR2IDX(qhp
->wq
.rq_wptr
,
283 qhp
->wq
.rq_size_log2
)].pbl_addr
= 0;
287 static int build_zero_stag_recv(struct iwch_qp
*qhp
, union t3_wr
*wqe
,
288 struct ib_recv_wr
*wr
)
296 * The T3 HW requires the PBL in the HW recv descriptor to reference
297 * a PBL entry. So we allocate the max needed PBL memory here and pass
298 * it to the uP in the recv WR. The uP will build the PBL and setup
299 * the HW recv descriptor.
301 pbl_addr
= cxio_hal_pblpool_alloc(&qhp
->rhp
->rdev
, T3_STAG0_PBL_SIZE
);
306 * Compute the 8B aligned offset.
308 pbl_offset
= (pbl_addr
- qhp
->rhp
->rdev
.rnic_info
.pbl_base
) >> 3;
310 wqe
->recv
.num_sgle
= cpu_to_be32(wr
->num_sge
);
312 for (i
= 0; i
< wr
->num_sge
; i
++) {
315 * Use a 128MB page size. This and an imposed 128MB
316 * sge length limit allows us to require only a 2-entry HW
317 * PBL for each SGE. This restriction is acceptable since
318 * since it is not possible to allocate 128MB of contiguous
319 * DMA coherent memory!
321 if (wr
->sg_list
[i
].length
> T3_STAG0_MAX_PBE_LEN
)
323 wqe
->recv
.pagesz
[i
] = T3_STAG0_PAGE_SHIFT
;
326 * T3 restricts a recv to all zero-stag or all non-zero-stag.
328 if (wr
->sg_list
[i
].lkey
!= 0)
330 wqe
->recv
.sgl
[i
].stag
= 0;
331 wqe
->recv
.sgl
[i
].len
= cpu_to_be32(wr
->sg_list
[i
].length
);
332 wqe
->recv
.sgl
[i
].to
= cpu_to_be64(wr
->sg_list
[i
].addr
);
333 wqe
->recv
.pbl_addr
[i
] = cpu_to_be32(pbl_offset
);
336 for (; i
< T3_MAX_SGE
; i
++) {
337 wqe
->recv
.pagesz
[i
] = 0;
338 wqe
->recv
.sgl
[i
].stag
= 0;
339 wqe
->recv
.sgl
[i
].len
= 0;
340 wqe
->recv
.sgl
[i
].to
= 0;
341 wqe
->recv
.pbl_addr
[i
] = 0;
343 qhp
->wq
.rq
[Q_PTR2IDX(qhp
->wq
.rq_wptr
,
344 qhp
->wq
.rq_size_log2
)].wr_id
= wr
->wr_id
;
345 qhp
->wq
.rq
[Q_PTR2IDX(qhp
->wq
.rq_wptr
,
346 qhp
->wq
.rq_size_log2
)].pbl_addr
= pbl_addr
;
350 int iwch_post_send(struct ib_qp
*ibqp
, struct ib_send_wr
*wr
,
351 struct ib_send_wr
**bad_wr
)
354 u8
uninitialized_var(t3_wr_flit_cnt
);
355 enum t3_wr_opcode t3_wr_opcode
= 0;
356 enum t3_wr_flags t3_wr_flags
;
365 qhp
= to_iwch_qp(ibqp
);
366 spin_lock_irqsave(&qhp
->lock
, flag
);
367 if (qhp
->attr
.state
> IWCH_QP_STATE_RTS
) {
368 spin_unlock_irqrestore(&qhp
->lock
, flag
);
372 num_wrs
= Q_FREECNT(qhp
->wq
.sq_rptr
, qhp
->wq
.sq_wptr
,
373 qhp
->wq
.sq_size_log2
);
375 spin_unlock_irqrestore(&qhp
->lock
, flag
);
384 idx
= Q_PTR2IDX(qhp
->wq
.wptr
, qhp
->wq
.size_log2
);
385 wqe
= (union t3_wr
*) (qhp
->wq
.queue
+ idx
);
387 if (wr
->send_flags
& IB_SEND_SOLICITED
)
388 t3_wr_flags
|= T3_SOLICITED_EVENT_FLAG
;
389 if (wr
->send_flags
& IB_SEND_SIGNALED
)
390 t3_wr_flags
|= T3_COMPLETION_FLAG
;
392 Q_PTR2IDX(qhp
->wq
.sq_wptr
, qhp
->wq
.sq_size_log2
);
393 switch (wr
->opcode
) {
395 case IB_WR_SEND_WITH_INV
:
396 if (wr
->send_flags
& IB_SEND_FENCE
)
397 t3_wr_flags
|= T3_READ_FENCE_FLAG
;
398 t3_wr_opcode
= T3_WR_SEND
;
399 err
= build_rdma_send(wqe
, wr
, &t3_wr_flit_cnt
);
401 case IB_WR_RDMA_WRITE
:
402 case IB_WR_RDMA_WRITE_WITH_IMM
:
403 t3_wr_opcode
= T3_WR_WRITE
;
404 err
= build_rdma_write(wqe
, wr
, &t3_wr_flit_cnt
);
406 case IB_WR_RDMA_READ
:
407 case IB_WR_RDMA_READ_WITH_INV
:
408 t3_wr_opcode
= T3_WR_READ
;
409 t3_wr_flags
= 0; /* T3 reads are always signaled */
410 err
= build_rdma_read(wqe
, wr
, &t3_wr_flit_cnt
);
413 sqp
->read_len
= wqe
->read
.local_len
;
414 if (!qhp
->wq
.oldest_read
)
415 qhp
->wq
.oldest_read
= sqp
;
417 case IB_WR_FAST_REG_MR
:
418 t3_wr_opcode
= T3_WR_FASTREG
;
419 err
= build_fastreg(wqe
, wr
, &t3_wr_flit_cnt
,
422 case IB_WR_LOCAL_INV
:
423 if (wr
->send_flags
& IB_SEND_FENCE
)
424 t3_wr_flags
|= T3_LOCAL_FENCE_FLAG
;
425 t3_wr_opcode
= T3_WR_INV_STAG
;
426 err
= build_inv_stag(wqe
, wr
, &t3_wr_flit_cnt
);
429 PDBG("%s post of type=%d TBD!\n", __func__
,
435 wqe
->send
.wrid
.id0
.hi
= qhp
->wq
.sq_wptr
;
436 sqp
->wr_id
= wr
->wr_id
;
437 sqp
->opcode
= wr2opcode(t3_wr_opcode
);
438 sqp
->sq_wptr
= qhp
->wq
.sq_wptr
;
440 sqp
->signaled
= (wr
->send_flags
& IB_SEND_SIGNALED
);
442 build_fw_riwrh((void *) wqe
, t3_wr_opcode
, t3_wr_flags
,
443 Q_GENBIT(qhp
->wq
.wptr
, qhp
->wq
.size_log2
),
445 (wr_cnt
== 1) ? T3_SOPEOP
: T3_SOP
);
446 PDBG("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
447 __func__
, (unsigned long long) wr
->wr_id
, idx
,
448 Q_PTR2IDX(qhp
->wq
.sq_wptr
, qhp
->wq
.sq_size_log2
),
452 qhp
->wq
.wptr
+= wr_cnt
;
455 spin_unlock_irqrestore(&qhp
->lock
, flag
);
456 if (cxio_wq_db_enabled(&qhp
->wq
))
457 ring_doorbell(qhp
->wq
.doorbell
, qhp
->wq
.qpid
);
465 int iwch_post_receive(struct ib_qp
*ibqp
, struct ib_recv_wr
*wr
,
466 struct ib_recv_wr
**bad_wr
)
475 qhp
= to_iwch_qp(ibqp
);
476 spin_lock_irqsave(&qhp
->lock
, flag
);
477 if (qhp
->attr
.state
> IWCH_QP_STATE_RTS
) {
478 spin_unlock_irqrestore(&qhp
->lock
, flag
);
482 num_wrs
= Q_FREECNT(qhp
->wq
.rq_rptr
, qhp
->wq
.rq_wptr
,
483 qhp
->wq
.rq_size_log2
) - 1;
485 spin_unlock_irqrestore(&qhp
->lock
, flag
);
490 if (wr
->num_sge
> T3_MAX_SGE
) {
494 idx
= Q_PTR2IDX(qhp
->wq
.wptr
, qhp
->wq
.size_log2
);
495 wqe
= (union t3_wr
*) (qhp
->wq
.queue
+ idx
);
497 if (wr
->sg_list
[0].lkey
)
498 err
= build_rdma_recv(qhp
, wqe
, wr
);
500 err
= build_zero_stag_recv(qhp
, wqe
, wr
);
507 build_fw_riwrh((void *) wqe
, T3_WR_RCV
, T3_COMPLETION_FLAG
,
508 Q_GENBIT(qhp
->wq
.wptr
, qhp
->wq
.size_log2
),
509 0, sizeof(struct t3_receive_wr
) >> 3, T3_SOPEOP
);
510 PDBG("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x "
511 "wqe %p \n", __func__
, (unsigned long long) wr
->wr_id
,
512 idx
, qhp
->wq
.rq_wptr
, qhp
->wq
.rq_rptr
, wqe
);
518 spin_unlock_irqrestore(&qhp
->lock
, flag
);
519 if (cxio_wq_db_enabled(&qhp
->wq
))
520 ring_doorbell(qhp
->wq
.doorbell
, qhp
->wq
.qpid
);
528 int iwch_bind_mw(struct ib_qp
*qp
,
530 struct ib_mw_bind
*mw_bind
)
532 struct iwch_dev
*rhp
;
542 enum t3_wr_flags t3_wr_flags
;
546 qhp
= to_iwch_qp(qp
);
547 mhp
= to_iwch_mw(mw
);
550 spin_lock_irqsave(&qhp
->lock
, flag
);
551 if (qhp
->attr
.state
> IWCH_QP_STATE_RTS
) {
552 spin_unlock_irqrestore(&qhp
->lock
, flag
);
555 num_wrs
= Q_FREECNT(qhp
->wq
.sq_rptr
, qhp
->wq
.sq_wptr
,
556 qhp
->wq
.sq_size_log2
);
558 spin_unlock_irqrestore(&qhp
->lock
, flag
);
561 idx
= Q_PTR2IDX(qhp
->wq
.wptr
, qhp
->wq
.size_log2
);
562 PDBG("%s: idx 0x%0x, mw 0x%p, mw_bind 0x%p\n", __func__
, idx
,
564 wqe
= (union t3_wr
*) (qhp
->wq
.queue
+ idx
);
567 if (mw_bind
->send_flags
& IB_SEND_SIGNALED
)
568 t3_wr_flags
= T3_COMPLETION_FLAG
;
570 sgl
.addr
= mw_bind
->addr
;
571 sgl
.lkey
= mw_bind
->mr
->lkey
;
572 sgl
.length
= mw_bind
->length
;
573 wqe
->bind
.reserved
= 0;
574 wqe
->bind
.type
= TPT_VATO
;
576 /* TBD: check perms */
577 wqe
->bind
.perms
= iwch_ib_to_tpt_bind_access(mw_bind
->mw_access_flags
);
578 wqe
->bind
.mr_stag
= cpu_to_be32(mw_bind
->mr
->lkey
);
579 wqe
->bind
.mw_stag
= cpu_to_be32(mw
->rkey
);
580 wqe
->bind
.mw_len
= cpu_to_be32(mw_bind
->length
);
581 wqe
->bind
.mw_va
= cpu_to_be64(mw_bind
->addr
);
582 err
= iwch_sgl2pbl_map(rhp
, &sgl
, 1, &pbl_addr
, &page_size
);
584 spin_unlock_irqrestore(&qhp
->lock
, flag
);
587 wqe
->send
.wrid
.id0
.hi
= qhp
->wq
.sq_wptr
;
588 sqp
= qhp
->wq
.sq
+ Q_PTR2IDX(qhp
->wq
.sq_wptr
, qhp
->wq
.sq_size_log2
);
589 sqp
->wr_id
= mw_bind
->wr_id
;
590 sqp
->opcode
= T3_BIND_MW
;
591 sqp
->sq_wptr
= qhp
->wq
.sq_wptr
;
593 sqp
->signaled
= (mw_bind
->send_flags
& IB_SEND_SIGNALED
);
594 wqe
->bind
.mr_pbl_addr
= cpu_to_be32(pbl_addr
);
595 wqe
->bind
.mr_pagesz
= page_size
;
596 build_fw_riwrh((void *)wqe
, T3_WR_BIND
, t3_wr_flags
,
597 Q_GENBIT(qhp
->wq
.wptr
, qhp
->wq
.size_log2
), 0,
598 sizeof(struct t3_bind_mw_wr
) >> 3, T3_SOPEOP
);
601 spin_unlock_irqrestore(&qhp
->lock
, flag
);
603 if (cxio_wq_db_enabled(&qhp
->wq
))
604 ring_doorbell(qhp
->wq
.doorbell
, qhp
->wq
.qpid
);
609 static inline void build_term_codes(struct respQ_msg_t
*rsp_msg
,
610 u8
*layer_type
, u8
*ecode
)
612 int status
= TPT_ERR_INTERNAL_ERR
;
619 status
= CQE_STATUS(rsp_msg
->cqe
);
620 opcode
= CQE_OPCODE(rsp_msg
->cqe
);
621 rqtype
= RQ_TYPE(rsp_msg
->cqe
);
622 send_inv
= (opcode
== T3_SEND_WITH_INV
) ||
623 (opcode
== T3_SEND_WITH_SE_INV
);
624 tagged
= (opcode
== T3_RDMA_WRITE
) ||
625 (rqtype
&& (opcode
== T3_READ_RESP
));
631 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_OP
;
632 *ecode
= RDMAP_CANT_INV_STAG
;
634 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
635 *ecode
= RDMAP_INV_STAG
;
639 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
640 if ((opcode
== T3_SEND_WITH_INV
) ||
641 (opcode
== T3_SEND_WITH_SE_INV
))
642 *ecode
= RDMAP_CANT_INV_STAG
;
644 *ecode
= RDMAP_STAG_NOT_ASSOC
;
647 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
648 *ecode
= RDMAP_STAG_NOT_ASSOC
;
651 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
652 *ecode
= RDMAP_ACC_VIOL
;
655 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
656 *ecode
= RDMAP_TO_WRAP
;
660 *layer_type
= LAYER_DDP
|DDP_TAGGED_ERR
;
661 *ecode
= DDPT_BASE_BOUNDS
;
663 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
664 *ecode
= RDMAP_BASE_BOUNDS
;
667 case TPT_ERR_INVALIDATE_SHARED_MR
:
668 case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND
:
669 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_OP
;
670 *ecode
= RDMAP_CANT_INV_STAG
;
673 case TPT_ERR_ECC_PSTAG
:
674 case TPT_ERR_INTERNAL_ERR
:
675 *layer_type
= LAYER_RDMAP
|RDMAP_LOCAL_CATA
;
678 case TPT_ERR_OUT_OF_RQE
:
679 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
680 *ecode
= DDPU_INV_MSN_NOBUF
;
682 case TPT_ERR_PBL_ADDR_BOUND
:
683 *layer_type
= LAYER_DDP
|DDP_TAGGED_ERR
;
684 *ecode
= DDPT_BASE_BOUNDS
;
687 *layer_type
= LAYER_MPA
|DDP_LLP
;
688 *ecode
= MPA_CRC_ERR
;
691 *layer_type
= LAYER_MPA
|DDP_LLP
;
692 *ecode
= MPA_MARKER_ERR
;
694 case TPT_ERR_PDU_LEN_ERR
:
695 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
696 *ecode
= DDPU_MSG_TOOBIG
;
698 case TPT_ERR_DDP_VERSION
:
700 *layer_type
= LAYER_DDP
|DDP_TAGGED_ERR
;
701 *ecode
= DDPT_INV_VERS
;
703 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
704 *ecode
= DDPU_INV_VERS
;
707 case TPT_ERR_RDMA_VERSION
:
708 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_OP
;
709 *ecode
= RDMAP_INV_VERS
;
712 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_OP
;
713 *ecode
= RDMAP_INV_OPCODE
;
715 case TPT_ERR_DDP_QUEUE_NUM
:
716 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
717 *ecode
= DDPU_INV_QN
;
720 case TPT_ERR_MSN_GAP
:
721 case TPT_ERR_MSN_RANGE
:
722 case TPT_ERR_IRD_OVERFLOW
:
723 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
724 *ecode
= DDPU_INV_MSN_RANGE
;
727 *layer_type
= LAYER_DDP
|DDP_LOCAL_CATA
;
731 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
732 *ecode
= DDPU_INV_MO
;
735 *layer_type
= LAYER_RDMAP
|DDP_LOCAL_CATA
;
741 int iwch_post_zb_read(struct iwch_qp
*qhp
)
745 u8 flit_cnt
= sizeof(struct t3_rdma_read_wr
) >> 3;
747 PDBG("%s enter\n", __func__
);
748 skb
= alloc_skb(40, GFP_KERNEL
);
750 printk(KERN_ERR
"%s cannot send zb_read!!\n", __func__
);
753 wqe
= (union t3_wr
*)skb_put(skb
, sizeof(struct t3_rdma_read_wr
));
754 memset(wqe
, 0, sizeof(struct t3_rdma_read_wr
));
755 wqe
->read
.rdmaop
= T3_READ_REQ
;
756 wqe
->read
.reserved
[0] = 0;
757 wqe
->read
.reserved
[1] = 0;
758 wqe
->read
.rem_stag
= cpu_to_be32(1);
759 wqe
->read
.rem_to
= cpu_to_be64(1);
760 wqe
->read
.local_stag
= cpu_to_be32(1);
761 wqe
->read
.local_len
= cpu_to_be32(0);
762 wqe
->read
.local_to
= cpu_to_be64(1);
763 wqe
->send
.wrh
.op_seop_flags
= cpu_to_be32(V_FW_RIWR_OP(T3_WR_READ
));
764 wqe
->send
.wrh
.gen_tid_len
= cpu_to_be32(V_FW_RIWR_TID(qhp
->ep
->hwtid
)|
765 V_FW_RIWR_LEN(flit_cnt
));
766 skb
->priority
= CPL_PRIORITY_DATA
;
767 return iwch_cxgb3_ofld_send(qhp
->rhp
->rdev
.t3cdev_p
, skb
);
771 * This posts a TERMINATE with layer=RDMA, type=catastrophic.
773 int iwch_post_terminate(struct iwch_qp
*qhp
, struct respQ_msg_t
*rsp_msg
)
776 struct terminate_message
*term
;
779 PDBG("%s %d\n", __func__
, __LINE__
);
780 skb
= alloc_skb(40, GFP_ATOMIC
);
782 printk(KERN_ERR
"%s cannot send TERMINATE!\n", __func__
);
785 wqe
= (union t3_wr
*)skb_put(skb
, 40);
787 wqe
->send
.rdmaop
= T3_TERMINATE
;
789 /* immediate data length */
790 wqe
->send
.plen
= htonl(4);
792 /* immediate data starts here. */
793 term
= (struct terminate_message
*)wqe
->send
.sgl
;
794 build_term_codes(rsp_msg
, &term
->layer_etype
, &term
->ecode
);
795 wqe
->send
.wrh
.op_seop_flags
= cpu_to_be32(V_FW_RIWR_OP(T3_WR_SEND
) |
796 V_FW_RIWR_FLAGS(T3_COMPLETION_FLAG
| T3_NOTIFY_FLAG
));
797 wqe
->send
.wrh
.gen_tid_len
= cpu_to_be32(V_FW_RIWR_TID(qhp
->ep
->hwtid
));
798 skb
->priority
= CPL_PRIORITY_DATA
;
799 return iwch_cxgb3_ofld_send(qhp
->rhp
->rdev
.t3cdev_p
, skb
);
803 * Assumes qhp lock is held.
805 static void __flush_qp(struct iwch_qp
*qhp
, struct iwch_cq
*rchp
,
806 struct iwch_cq
*schp
, unsigned long *flag
)
812 PDBG("%s qhp %p rchp %p schp %p\n", __func__
, qhp
, rchp
, schp
);
813 /* take a ref on the qhp since we must release the lock */
814 atomic_inc(&qhp
->refcnt
);
815 spin_unlock_irqrestore(&qhp
->lock
, *flag
);
817 /* locking hierarchy: cq lock first, then qp lock. */
818 spin_lock_irqsave(&rchp
->lock
, *flag
);
819 spin_lock(&qhp
->lock
);
820 cxio_flush_hw_cq(&rchp
->cq
);
821 cxio_count_rcqes(&rchp
->cq
, &qhp
->wq
, &count
);
822 flushed
= cxio_flush_rq(&qhp
->wq
, &rchp
->cq
, count
);
823 spin_unlock(&qhp
->lock
);
824 spin_unlock_irqrestore(&rchp
->lock
, *flag
);
826 (*rchp
->ibcq
.comp_handler
)(&rchp
->ibcq
, rchp
->ibcq
.cq_context
);
828 /* locking hierarchy: cq lock first, then qp lock. */
829 spin_lock_irqsave(&schp
->lock
, *flag
);
830 spin_lock(&qhp
->lock
);
831 cxio_flush_hw_cq(&schp
->cq
);
832 cxio_count_scqes(&schp
->cq
, &qhp
->wq
, &count
);
833 flushed
= cxio_flush_sq(&qhp
->wq
, &schp
->cq
, count
);
834 spin_unlock(&qhp
->lock
);
835 spin_unlock_irqrestore(&schp
->lock
, *flag
);
837 (*schp
->ibcq
.comp_handler
)(&schp
->ibcq
, schp
->ibcq
.cq_context
);
840 if (atomic_dec_and_test(&qhp
->refcnt
))
843 spin_lock_irqsave(&qhp
->lock
, *flag
);
846 static void flush_qp(struct iwch_qp
*qhp
, unsigned long *flag
)
848 struct iwch_cq
*rchp
, *schp
;
850 rchp
= get_chp(qhp
->rhp
, qhp
->attr
.rcq
);
851 schp
= get_chp(qhp
->rhp
, qhp
->attr
.scq
);
853 if (qhp
->ibqp
.uobject
) {
854 cxio_set_wq_in_error(&qhp
->wq
);
855 cxio_set_cq_in_error(&rchp
->cq
);
856 (*rchp
->ibcq
.comp_handler
)(&rchp
->ibcq
, rchp
->ibcq
.cq_context
);
858 cxio_set_cq_in_error(&schp
->cq
);
859 (*schp
->ibcq
.comp_handler
)(&schp
->ibcq
,
860 schp
->ibcq
.cq_context
);
864 __flush_qp(qhp
, rchp
, schp
, flag
);
869 * Return count of RECV WRs posted
871 u16
iwch_rqes_posted(struct iwch_qp
*qhp
)
873 union t3_wr
*wqe
= qhp
->wq
.queue
;
875 while ((count
+1) != 0 && fw_riwrh_opcode((struct fw_riwrh
*)wqe
) == T3_WR_RCV
) {
879 PDBG("%s qhp %p count %u\n", __func__
, qhp
, count
);
883 static int rdma_init(struct iwch_dev
*rhp
, struct iwch_qp
*qhp
,
884 enum iwch_qp_attr_mask mask
,
885 struct iwch_qp_attributes
*attrs
)
887 struct t3_rdma_init_attr init_attr
;
890 init_attr
.tid
= qhp
->ep
->hwtid
;
891 init_attr
.qpid
= qhp
->wq
.qpid
;
892 init_attr
.pdid
= qhp
->attr
.pd
;
893 init_attr
.scqid
= qhp
->attr
.scq
;
894 init_attr
.rcqid
= qhp
->attr
.rcq
;
895 init_attr
.rq_addr
= qhp
->wq
.rq_addr
;
896 init_attr
.rq_size
= 1 << qhp
->wq
.rq_size_log2
;
897 init_attr
.mpaattrs
= uP_RI_MPA_IETF_ENABLE
|
898 qhp
->attr
.mpa_attr
.recv_marker_enabled
|
899 (qhp
->attr
.mpa_attr
.xmit_marker_enabled
<< 1) |
900 (qhp
->attr
.mpa_attr
.crc_enabled
<< 2);
902 init_attr
.qpcaps
= uP_RI_QP_RDMA_READ_ENABLE
|
903 uP_RI_QP_RDMA_WRITE_ENABLE
|
904 uP_RI_QP_BIND_ENABLE
;
905 if (!qhp
->ibqp
.uobject
)
906 init_attr
.qpcaps
|= uP_RI_QP_STAG0_ENABLE
|
907 uP_RI_QP_FAST_REGISTER_ENABLE
;
909 init_attr
.tcp_emss
= qhp
->ep
->emss
;
910 init_attr
.ord
= qhp
->attr
.max_ord
;
911 init_attr
.ird
= qhp
->attr
.max_ird
;
912 init_attr
.qp_dma_addr
= qhp
->wq
.dma_addr
;
913 init_attr
.qp_dma_size
= (1UL << qhp
->wq
.size_log2
);
914 init_attr
.rqe_count
= iwch_rqes_posted(qhp
);
915 init_attr
.flags
= qhp
->attr
.mpa_attr
.initiator
? MPA_INITIATOR
: 0;
916 init_attr
.chan
= qhp
->ep
->l2t
->smt_idx
;
918 init_attr
.rtr_type
= RTR_READ
;
919 if (init_attr
.ord
== 0 && qhp
->attr
.mpa_attr
.initiator
)
921 if (init_attr
.ird
== 0 && !qhp
->attr
.mpa_attr
.initiator
)
924 init_attr
.rtr_type
= 0;
925 init_attr
.irs
= qhp
->ep
->rcv_seq
;
926 PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
927 "flags 0x%x qpcaps 0x%x\n", __func__
,
928 init_attr
.rq_addr
, init_attr
.rq_size
,
929 init_attr
.flags
, init_attr
.qpcaps
);
930 ret
= cxio_rdma_init(&rhp
->rdev
, &init_attr
);
931 PDBG("%s ret %d\n", __func__
, ret
);
935 int iwch_modify_qp(struct iwch_dev
*rhp
, struct iwch_qp
*qhp
,
936 enum iwch_qp_attr_mask mask
,
937 struct iwch_qp_attributes
*attrs
,
941 struct iwch_qp_attributes newattr
= qhp
->attr
;
947 struct iwch_ep
*ep
= NULL
;
949 PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __func__
,
950 qhp
, qhp
->wq
.qpid
, qhp
->ep
, qhp
->attr
.state
,
951 (mask
& IWCH_QP_ATTR_NEXT_STATE
) ? attrs
->next_state
: -1);
953 spin_lock_irqsave(&qhp
->lock
, flag
);
955 /* Process attr changes if in IDLE */
956 if (mask
& IWCH_QP_ATTR_VALID_MODIFY
) {
957 if (qhp
->attr
.state
!= IWCH_QP_STATE_IDLE
) {
961 if (mask
& IWCH_QP_ATTR_ENABLE_RDMA_READ
)
962 newattr
.enable_rdma_read
= attrs
->enable_rdma_read
;
963 if (mask
& IWCH_QP_ATTR_ENABLE_RDMA_WRITE
)
964 newattr
.enable_rdma_write
= attrs
->enable_rdma_write
;
965 if (mask
& IWCH_QP_ATTR_ENABLE_RDMA_BIND
)
966 newattr
.enable_bind
= attrs
->enable_bind
;
967 if (mask
& IWCH_QP_ATTR_MAX_ORD
) {
969 rhp
->attr
.max_rdma_read_qp_depth
) {
973 newattr
.max_ord
= attrs
->max_ord
;
975 if (mask
& IWCH_QP_ATTR_MAX_IRD
) {
977 rhp
->attr
.max_rdma_reads_per_qp
) {
981 newattr
.max_ird
= attrs
->max_ird
;
986 if (!(mask
& IWCH_QP_ATTR_NEXT_STATE
))
988 if (qhp
->attr
.state
== attrs
->next_state
)
991 switch (qhp
->attr
.state
) {
992 case IWCH_QP_STATE_IDLE
:
993 switch (attrs
->next_state
) {
994 case IWCH_QP_STATE_RTS
:
995 if (!(mask
& IWCH_QP_ATTR_LLP_STREAM_HANDLE
)) {
999 if (!(mask
& IWCH_QP_ATTR_MPA_ATTR
)) {
1003 qhp
->attr
.mpa_attr
= attrs
->mpa_attr
;
1004 qhp
->attr
.llp_stream_handle
= attrs
->llp_stream_handle
;
1005 qhp
->ep
= qhp
->attr
.llp_stream_handle
;
1006 qhp
->attr
.state
= IWCH_QP_STATE_RTS
;
1009 * Ref the endpoint here and deref when we
1010 * disassociate the endpoint from the QP. This
1011 * happens in CLOSING->IDLE transition or *->ERROR
1014 get_ep(&qhp
->ep
->com
);
1015 spin_unlock_irqrestore(&qhp
->lock
, flag
);
1016 ret
= rdma_init(rhp
, qhp
, mask
, attrs
);
1017 spin_lock_irqsave(&qhp
->lock
, flag
);
1021 case IWCH_QP_STATE_ERROR
:
1022 qhp
->attr
.state
= IWCH_QP_STATE_ERROR
;
1023 flush_qp(qhp
, &flag
);
1030 case IWCH_QP_STATE_RTS
:
1031 switch (attrs
->next_state
) {
1032 case IWCH_QP_STATE_CLOSING
:
1033 BUG_ON(atomic_read(&qhp
->ep
->com
.kref
.refcount
) < 2);
1034 qhp
->attr
.state
= IWCH_QP_STATE_CLOSING
;
1042 case IWCH_QP_STATE_TERMINATE
:
1043 qhp
->attr
.state
= IWCH_QP_STATE_TERMINATE
;
1044 if (qhp
->ibqp
.uobject
)
1045 cxio_set_wq_in_error(&qhp
->wq
);
1049 case IWCH_QP_STATE_ERROR
:
1050 qhp
->attr
.state
= IWCH_QP_STATE_ERROR
;
1064 case IWCH_QP_STATE_CLOSING
:
1069 switch (attrs
->next_state
) {
1070 case IWCH_QP_STATE_IDLE
:
1071 flush_qp(qhp
, &flag
);
1072 qhp
->attr
.state
= IWCH_QP_STATE_IDLE
;
1073 qhp
->attr
.llp_stream_handle
= NULL
;
1074 put_ep(&qhp
->ep
->com
);
1076 wake_up(&qhp
->wait
);
1078 case IWCH_QP_STATE_ERROR
:
1085 case IWCH_QP_STATE_ERROR
:
1086 if (attrs
->next_state
!= IWCH_QP_STATE_IDLE
) {
1091 if (!Q_EMPTY(qhp
->wq
.sq_rptr
, qhp
->wq
.sq_wptr
) ||
1092 !Q_EMPTY(qhp
->wq
.rq_rptr
, qhp
->wq
.rq_wptr
)) {
1096 qhp
->attr
.state
= IWCH_QP_STATE_IDLE
;
1098 case IWCH_QP_STATE_TERMINATE
:
1106 printk(KERN_ERR
"%s in a bad state %d\n",
1107 __func__
, qhp
->attr
.state
);
1114 PDBG("%s disassociating ep %p qpid 0x%x\n", __func__
, qhp
->ep
,
1117 /* disassociate the LLP connection */
1118 qhp
->attr
.llp_stream_handle
= NULL
;
1121 qhp
->attr
.state
= IWCH_QP_STATE_ERROR
;
1123 wake_up(&qhp
->wait
);
1125 flush_qp(qhp
, &flag
);
1127 spin_unlock_irqrestore(&qhp
->lock
, flag
);
1130 iwch_post_terminate(qhp
, NULL
);
1133 * If disconnect is 1, then we need to initiate a disconnect
1134 * on the EP. This can be a normal close (RTS->CLOSING) or
1135 * an abnormal close (RTS/CLOSING->ERROR).
1138 iwch_ep_disconnect(ep
, abort
, GFP_KERNEL
);
1143 * If free is 1, then we've disassociated the EP from the QP
1144 * and we need to dereference the EP.
1149 PDBG("%s exit state %d\n", __func__
, qhp
->attr
.state
);
1153 static int quiesce_qp(struct iwch_qp
*qhp
)
1155 spin_lock_irq(&qhp
->lock
);
1156 iwch_quiesce_tid(qhp
->ep
);
1157 qhp
->flags
|= QP_QUIESCED
;
1158 spin_unlock_irq(&qhp
->lock
);
1162 static int resume_qp(struct iwch_qp
*qhp
)
1164 spin_lock_irq(&qhp
->lock
);
1165 iwch_resume_tid(qhp
->ep
);
1166 qhp
->flags
&= ~QP_QUIESCED
;
1167 spin_unlock_irq(&qhp
->lock
);
1171 int iwch_quiesce_qps(struct iwch_cq
*chp
)
1174 struct iwch_qp
*qhp
;
1176 for (i
=0; i
< T3_MAX_NUM_QP
; i
++) {
1177 qhp
= get_qhp(chp
->rhp
, i
);
1180 if ((qhp
->attr
.rcq
== chp
->cq
.cqid
) && !qp_quiesced(qhp
)) {
1184 if ((qhp
->attr
.scq
== chp
->cq
.cqid
) && !qp_quiesced(qhp
))
1190 int iwch_resume_qps(struct iwch_cq
*chp
)
1193 struct iwch_qp
*qhp
;
1195 for (i
=0; i
< T3_MAX_NUM_QP
; i
++) {
1196 qhp
= get_qhp(chp
->rhp
, i
);
1199 if ((qhp
->attr
.rcq
== chp
->cq
.cqid
) && qp_quiesced(qhp
)) {
1203 if ((qhp
->attr
.scq
== chp
->cq
.cqid
) && qp_quiesced(qhp
))