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(rdma_wr(wr
)->rkey
);
99 wqe
->write
.to_sink
= cpu_to_be64(rdma_wr(wr
)->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(rdma_wr(wr
)->rkey
);
141 wqe
->read
.rem_to
= cpu_to_be64(rdma_wr(wr
)->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_memreg(union t3_wr
*wqe
, struct ib_reg_wr
*wr
,
150 u8
*flit_cnt
, int *wr_cnt
, struct t3_wq
*wq
)
152 struct iwch_mr
*mhp
= to_iwch_mr(wr
->mr
);
156 if (mhp
->npages
> T3_MAX_FASTREG_DEPTH
)
159 wqe
->fastreg
.stag
= cpu_to_be32(wr
->key
);
160 wqe
->fastreg
.len
= cpu_to_be32(mhp
->ibmr
.length
);
161 wqe
->fastreg
.va_base_hi
= cpu_to_be32(mhp
->ibmr
.iova
>> 32);
162 wqe
->fastreg
.va_base_lo_fbo
=
163 cpu_to_be32(mhp
->ibmr
.iova
& 0xffffffff);
164 wqe
->fastreg
.page_type_perms
= cpu_to_be32(
165 V_FR_PAGE_COUNT(mhp
->npages
) |
166 V_FR_PAGE_SIZE(ilog2(wr
->mr
->page_size
) - 12) |
167 V_FR_TYPE(TPT_VATO
) |
168 V_FR_PERMS(iwch_ib_to_tpt_access(wr
->access
)));
169 p
= &wqe
->fastreg
.pbl_addrs
[0];
170 for (i
= 0; i
< mhp
->npages
; i
++, p
++) {
172 /* If we need a 2nd WR, then set it up */
173 if (i
== T3_MAX_FASTREG_FRAG
) {
175 wqe
= (union t3_wr
*)(wq
->queue
+
176 Q_PTR2IDX((wq
->wptr
+1), wq
->size_log2
));
177 build_fw_riwrh((void *)wqe
, T3_WR_FASTREG
, 0,
178 Q_GENBIT(wq
->wptr
+ 1, wq
->size_log2
),
179 0, 1 + mhp
->npages
- T3_MAX_FASTREG_FRAG
,
182 p
= &wqe
->pbl_frag
.pbl_addrs
[0];
184 *p
= cpu_to_be64((u64
)mhp
->pages
[i
]);
186 *flit_cnt
= 5 + mhp
->npages
;
192 static int build_inv_stag(union t3_wr
*wqe
, struct ib_send_wr
*wr
,
195 wqe
->local_inv
.stag
= cpu_to_be32(wr
->ex
.invalidate_rkey
);
196 wqe
->local_inv
.reserved
= 0;
197 *flit_cnt
= sizeof(struct t3_local_inv_wr
) >> 3;
201 static int iwch_sgl2pbl_map(struct iwch_dev
*rhp
, struct ib_sge
*sg_list
,
202 u32 num_sgle
, u32
* pbl_addr
, u8
* page_size
)
207 for (i
= 0; i
< num_sgle
; i
++) {
209 mhp
= get_mhp(rhp
, (sg_list
[i
].lkey
) >> 8);
211 PDBG("%s %d\n", __func__
, __LINE__
);
214 if (!mhp
->attr
.state
) {
215 PDBG("%s %d\n", __func__
, __LINE__
);
218 if (mhp
->attr
.zbva
) {
219 PDBG("%s %d\n", __func__
, __LINE__
);
223 if (sg_list
[i
].addr
< mhp
->attr
.va_fbo
) {
224 PDBG("%s %d\n", __func__
, __LINE__
);
227 if (sg_list
[i
].addr
+ ((u64
) sg_list
[i
].length
) <
229 PDBG("%s %d\n", __func__
, __LINE__
);
232 if (sg_list
[i
].addr
+ ((u64
) sg_list
[i
].length
) >
233 mhp
->attr
.va_fbo
+ ((u64
) mhp
->attr
.len
)) {
234 PDBG("%s %d\n", __func__
, __LINE__
);
237 offset
= sg_list
[i
].addr
- mhp
->attr
.va_fbo
;
238 offset
+= mhp
->attr
.va_fbo
&
239 ((1UL << (12 + mhp
->attr
.page_size
)) - 1);
240 pbl_addr
[i
] = ((mhp
->attr
.pbl_addr
-
241 rhp
->rdev
.rnic_info
.pbl_base
) >> 3) +
242 (offset
>> (12 + mhp
->attr
.page_size
));
243 page_size
[i
] = mhp
->attr
.page_size
;
248 static int build_rdma_recv(struct iwch_qp
*qhp
, union t3_wr
*wqe
,
249 struct ib_recv_wr
*wr
)
252 u32 pbl_addr
[T3_MAX_SGE
];
253 u8 page_size
[T3_MAX_SGE
];
255 err
= iwch_sgl2pbl_map(qhp
->rhp
, wr
->sg_list
, wr
->num_sge
, pbl_addr
,
259 wqe
->recv
.pagesz
[0] = page_size
[0];
260 wqe
->recv
.pagesz
[1] = page_size
[1];
261 wqe
->recv
.pagesz
[2] = page_size
[2];
262 wqe
->recv
.pagesz
[3] = page_size
[3];
263 wqe
->recv
.num_sgle
= cpu_to_be32(wr
->num_sge
);
264 for (i
= 0; i
< wr
->num_sge
; i
++) {
265 wqe
->recv
.sgl
[i
].stag
= cpu_to_be32(wr
->sg_list
[i
].lkey
);
266 wqe
->recv
.sgl
[i
].len
= cpu_to_be32(wr
->sg_list
[i
].length
);
268 /* to in the WQE == the offset into the page */
269 wqe
->recv
.sgl
[i
].to
= cpu_to_be64(((u32
)wr
->sg_list
[i
].addr
) &
270 ((1UL << (12 + page_size
[i
])) - 1));
272 /* pbl_addr is the adapters address in the PBL */
273 wqe
->recv
.pbl_addr
[i
] = cpu_to_be32(pbl_addr
[i
]);
275 for (; i
< T3_MAX_SGE
; i
++) {
276 wqe
->recv
.sgl
[i
].stag
= 0;
277 wqe
->recv
.sgl
[i
].len
= 0;
278 wqe
->recv
.sgl
[i
].to
= 0;
279 wqe
->recv
.pbl_addr
[i
] = 0;
281 qhp
->wq
.rq
[Q_PTR2IDX(qhp
->wq
.rq_wptr
,
282 qhp
->wq
.rq_size_log2
)].wr_id
= wr
->wr_id
;
283 qhp
->wq
.rq
[Q_PTR2IDX(qhp
->wq
.rq_wptr
,
284 qhp
->wq
.rq_size_log2
)].pbl_addr
= 0;
288 static int build_zero_stag_recv(struct iwch_qp
*qhp
, union t3_wr
*wqe
,
289 struct ib_recv_wr
*wr
)
297 * The T3 HW requires the PBL in the HW recv descriptor to reference
298 * a PBL entry. So we allocate the max needed PBL memory here and pass
299 * it to the uP in the recv WR. The uP will build the PBL and setup
300 * the HW recv descriptor.
302 pbl_addr
= cxio_hal_pblpool_alloc(&qhp
->rhp
->rdev
, T3_STAG0_PBL_SIZE
);
307 * Compute the 8B aligned offset.
309 pbl_offset
= (pbl_addr
- qhp
->rhp
->rdev
.rnic_info
.pbl_base
) >> 3;
311 wqe
->recv
.num_sgle
= cpu_to_be32(wr
->num_sge
);
313 for (i
= 0; i
< wr
->num_sge
; i
++) {
316 * Use a 128MB page size. This and an imposed 128MB
317 * sge length limit allows us to require only a 2-entry HW
318 * PBL for each SGE. This restriction is acceptable since
319 * since it is not possible to allocate 128MB of contiguous
320 * DMA coherent memory!
322 if (wr
->sg_list
[i
].length
> T3_STAG0_MAX_PBE_LEN
)
324 wqe
->recv
.pagesz
[i
] = T3_STAG0_PAGE_SHIFT
;
327 * T3 restricts a recv to all zero-stag or all non-zero-stag.
329 if (wr
->sg_list
[i
].lkey
!= 0)
331 wqe
->recv
.sgl
[i
].stag
= 0;
332 wqe
->recv
.sgl
[i
].len
= cpu_to_be32(wr
->sg_list
[i
].length
);
333 wqe
->recv
.sgl
[i
].to
= cpu_to_be64(wr
->sg_list
[i
].addr
);
334 wqe
->recv
.pbl_addr
[i
] = cpu_to_be32(pbl_offset
);
337 for (; i
< T3_MAX_SGE
; i
++) {
338 wqe
->recv
.pagesz
[i
] = 0;
339 wqe
->recv
.sgl
[i
].stag
= 0;
340 wqe
->recv
.sgl
[i
].len
= 0;
341 wqe
->recv
.sgl
[i
].to
= 0;
342 wqe
->recv
.pbl_addr
[i
] = 0;
344 qhp
->wq
.rq
[Q_PTR2IDX(qhp
->wq
.rq_wptr
,
345 qhp
->wq
.rq_size_log2
)].wr_id
= wr
->wr_id
;
346 qhp
->wq
.rq
[Q_PTR2IDX(qhp
->wq
.rq_wptr
,
347 qhp
->wq
.rq_size_log2
)].pbl_addr
= pbl_addr
;
351 int iwch_post_send(struct ib_qp
*ibqp
, struct ib_send_wr
*wr
,
352 struct ib_send_wr
**bad_wr
)
355 u8
uninitialized_var(t3_wr_flit_cnt
);
356 enum t3_wr_opcode t3_wr_opcode
= 0;
357 enum t3_wr_flags t3_wr_flags
;
366 qhp
= to_iwch_qp(ibqp
);
367 spin_lock_irqsave(&qhp
->lock
, flag
);
368 if (qhp
->attr
.state
> IWCH_QP_STATE_RTS
) {
369 spin_unlock_irqrestore(&qhp
->lock
, flag
);
373 num_wrs
= Q_FREECNT(qhp
->wq
.sq_rptr
, qhp
->wq
.sq_wptr
,
374 qhp
->wq
.sq_size_log2
);
376 spin_unlock_irqrestore(&qhp
->lock
, flag
);
385 idx
= Q_PTR2IDX(qhp
->wq
.wptr
, qhp
->wq
.size_log2
);
386 wqe
= (union t3_wr
*) (qhp
->wq
.queue
+ idx
);
388 if (wr
->send_flags
& IB_SEND_SOLICITED
)
389 t3_wr_flags
|= T3_SOLICITED_EVENT_FLAG
;
390 if (wr
->send_flags
& IB_SEND_SIGNALED
)
391 t3_wr_flags
|= T3_COMPLETION_FLAG
;
393 Q_PTR2IDX(qhp
->wq
.sq_wptr
, qhp
->wq
.sq_size_log2
);
394 switch (wr
->opcode
) {
396 case IB_WR_SEND_WITH_INV
:
397 if (wr
->send_flags
& IB_SEND_FENCE
)
398 t3_wr_flags
|= T3_READ_FENCE_FLAG
;
399 t3_wr_opcode
= T3_WR_SEND
;
400 err
= build_rdma_send(wqe
, wr
, &t3_wr_flit_cnt
);
402 case IB_WR_RDMA_WRITE
:
403 case IB_WR_RDMA_WRITE_WITH_IMM
:
404 t3_wr_opcode
= T3_WR_WRITE
;
405 err
= build_rdma_write(wqe
, wr
, &t3_wr_flit_cnt
);
407 case IB_WR_RDMA_READ
:
408 case IB_WR_RDMA_READ_WITH_INV
:
409 t3_wr_opcode
= T3_WR_READ
;
410 t3_wr_flags
= 0; /* T3 reads are always signaled */
411 err
= build_rdma_read(wqe
, wr
, &t3_wr_flit_cnt
);
414 sqp
->read_len
= wqe
->read
.local_len
;
415 if (!qhp
->wq
.oldest_read
)
416 qhp
->wq
.oldest_read
= sqp
;
419 t3_wr_opcode
= T3_WR_FASTREG
;
420 err
= build_memreg(wqe
, reg_wr(wr
), &t3_wr_flit_cnt
,
423 case IB_WR_LOCAL_INV
:
424 if (wr
->send_flags
& IB_SEND_FENCE
)
425 t3_wr_flags
|= T3_LOCAL_FENCE_FLAG
;
426 t3_wr_opcode
= T3_WR_INV_STAG
;
427 err
= build_inv_stag(wqe
, wr
, &t3_wr_flit_cnt
);
430 PDBG("%s post of type=%d TBD!\n", __func__
,
436 wqe
->send
.wrid
.id0
.hi
= qhp
->wq
.sq_wptr
;
437 sqp
->wr_id
= wr
->wr_id
;
438 sqp
->opcode
= wr2opcode(t3_wr_opcode
);
439 sqp
->sq_wptr
= qhp
->wq
.sq_wptr
;
441 sqp
->signaled
= (wr
->send_flags
& IB_SEND_SIGNALED
);
443 build_fw_riwrh((void *) wqe
, t3_wr_opcode
, t3_wr_flags
,
444 Q_GENBIT(qhp
->wq
.wptr
, qhp
->wq
.size_log2
),
446 (wr_cnt
== 1) ? T3_SOPEOP
: T3_SOP
);
447 PDBG("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
448 __func__
, (unsigned long long) wr
->wr_id
, idx
,
449 Q_PTR2IDX(qhp
->wq
.sq_wptr
, qhp
->wq
.sq_size_log2
),
453 qhp
->wq
.wptr
+= wr_cnt
;
456 spin_unlock_irqrestore(&qhp
->lock
, flag
);
457 if (cxio_wq_db_enabled(&qhp
->wq
))
458 ring_doorbell(qhp
->wq
.doorbell
, qhp
->wq
.qpid
);
466 int iwch_post_receive(struct ib_qp
*ibqp
, struct ib_recv_wr
*wr
,
467 struct ib_recv_wr
**bad_wr
)
476 qhp
= to_iwch_qp(ibqp
);
477 spin_lock_irqsave(&qhp
->lock
, flag
);
478 if (qhp
->attr
.state
> IWCH_QP_STATE_RTS
) {
479 spin_unlock_irqrestore(&qhp
->lock
, flag
);
483 num_wrs
= Q_FREECNT(qhp
->wq
.rq_rptr
, qhp
->wq
.rq_wptr
,
484 qhp
->wq
.rq_size_log2
) - 1;
486 spin_unlock_irqrestore(&qhp
->lock
, flag
);
491 if (wr
->num_sge
> T3_MAX_SGE
) {
495 idx
= Q_PTR2IDX(qhp
->wq
.wptr
, qhp
->wq
.size_log2
);
496 wqe
= (union t3_wr
*) (qhp
->wq
.queue
+ idx
);
498 if (wr
->sg_list
[0].lkey
)
499 err
= build_rdma_recv(qhp
, wqe
, wr
);
501 err
= build_zero_stag_recv(qhp
, wqe
, wr
);
508 build_fw_riwrh((void *) wqe
, T3_WR_RCV
, T3_COMPLETION_FLAG
,
509 Q_GENBIT(qhp
->wq
.wptr
, qhp
->wq
.size_log2
),
510 0, sizeof(struct t3_receive_wr
) >> 3, T3_SOPEOP
);
511 PDBG("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x "
512 "wqe %p \n", __func__
, (unsigned long long) wr
->wr_id
,
513 idx
, qhp
->wq
.rq_wptr
, qhp
->wq
.rq_rptr
, wqe
);
519 spin_unlock_irqrestore(&qhp
->lock
, flag
);
520 if (cxio_wq_db_enabled(&qhp
->wq
))
521 ring_doorbell(qhp
->wq
.doorbell
, qhp
->wq
.qpid
);
529 int iwch_bind_mw(struct ib_qp
*qp
,
531 struct ib_mw_bind
*mw_bind
)
533 struct iwch_dev
*rhp
;
543 enum t3_wr_flags t3_wr_flags
;
547 qhp
= to_iwch_qp(qp
);
548 mhp
= to_iwch_mw(mw
);
551 spin_lock_irqsave(&qhp
->lock
, flag
);
552 if (qhp
->attr
.state
> IWCH_QP_STATE_RTS
) {
553 spin_unlock_irqrestore(&qhp
->lock
, flag
);
556 num_wrs
= Q_FREECNT(qhp
->wq
.sq_rptr
, qhp
->wq
.sq_wptr
,
557 qhp
->wq
.sq_size_log2
);
559 spin_unlock_irqrestore(&qhp
->lock
, flag
);
562 idx
= Q_PTR2IDX(qhp
->wq
.wptr
, qhp
->wq
.size_log2
);
563 PDBG("%s: idx 0x%0x, mw 0x%p, mw_bind 0x%p\n", __func__
, idx
,
565 wqe
= (union t3_wr
*) (qhp
->wq
.queue
+ idx
);
568 if (mw_bind
->send_flags
& IB_SEND_SIGNALED
)
569 t3_wr_flags
= T3_COMPLETION_FLAG
;
571 sgl
.addr
= mw_bind
->bind_info
.addr
;
572 sgl
.lkey
= mw_bind
->bind_info
.mr
->lkey
;
573 sgl
.length
= mw_bind
->bind_info
.length
;
574 wqe
->bind
.reserved
= 0;
575 wqe
->bind
.type
= TPT_VATO
;
577 /* TBD: check perms */
578 wqe
->bind
.perms
= iwch_ib_to_tpt_bind_access(
579 mw_bind
->bind_info
.mw_access_flags
);
580 wqe
->bind
.mr_stag
= cpu_to_be32(mw_bind
->bind_info
.mr
->lkey
);
581 wqe
->bind
.mw_stag
= cpu_to_be32(mw
->rkey
);
582 wqe
->bind
.mw_len
= cpu_to_be32(mw_bind
->bind_info
.length
);
583 wqe
->bind
.mw_va
= cpu_to_be64(mw_bind
->bind_info
.addr
);
584 err
= iwch_sgl2pbl_map(rhp
, &sgl
, 1, &pbl_addr
, &page_size
);
586 spin_unlock_irqrestore(&qhp
->lock
, flag
);
589 wqe
->send
.wrid
.id0
.hi
= qhp
->wq
.sq_wptr
;
590 sqp
= qhp
->wq
.sq
+ Q_PTR2IDX(qhp
->wq
.sq_wptr
, qhp
->wq
.sq_size_log2
);
591 sqp
->wr_id
= mw_bind
->wr_id
;
592 sqp
->opcode
= T3_BIND_MW
;
593 sqp
->sq_wptr
= qhp
->wq
.sq_wptr
;
595 sqp
->signaled
= (mw_bind
->send_flags
& IB_SEND_SIGNALED
);
596 wqe
->bind
.mr_pbl_addr
= cpu_to_be32(pbl_addr
);
597 wqe
->bind
.mr_pagesz
= page_size
;
598 build_fw_riwrh((void *)wqe
, T3_WR_BIND
, t3_wr_flags
,
599 Q_GENBIT(qhp
->wq
.wptr
, qhp
->wq
.size_log2
), 0,
600 sizeof(struct t3_bind_mw_wr
) >> 3, T3_SOPEOP
);
603 spin_unlock_irqrestore(&qhp
->lock
, flag
);
605 if (cxio_wq_db_enabled(&qhp
->wq
))
606 ring_doorbell(qhp
->wq
.doorbell
, qhp
->wq
.qpid
);
611 static inline void build_term_codes(struct respQ_msg_t
*rsp_msg
,
612 u8
*layer_type
, u8
*ecode
)
614 int status
= TPT_ERR_INTERNAL_ERR
;
621 status
= CQE_STATUS(rsp_msg
->cqe
);
622 opcode
= CQE_OPCODE(rsp_msg
->cqe
);
623 rqtype
= RQ_TYPE(rsp_msg
->cqe
);
624 send_inv
= (opcode
== T3_SEND_WITH_INV
) ||
625 (opcode
== T3_SEND_WITH_SE_INV
);
626 tagged
= (opcode
== T3_RDMA_WRITE
) ||
627 (rqtype
&& (opcode
== T3_READ_RESP
));
633 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_OP
;
634 *ecode
= RDMAP_CANT_INV_STAG
;
636 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
637 *ecode
= RDMAP_INV_STAG
;
641 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
642 if ((opcode
== T3_SEND_WITH_INV
) ||
643 (opcode
== T3_SEND_WITH_SE_INV
))
644 *ecode
= RDMAP_CANT_INV_STAG
;
646 *ecode
= RDMAP_STAG_NOT_ASSOC
;
649 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
650 *ecode
= RDMAP_STAG_NOT_ASSOC
;
653 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
654 *ecode
= RDMAP_ACC_VIOL
;
657 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
658 *ecode
= RDMAP_TO_WRAP
;
662 *layer_type
= LAYER_DDP
|DDP_TAGGED_ERR
;
663 *ecode
= DDPT_BASE_BOUNDS
;
665 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_PROT
;
666 *ecode
= RDMAP_BASE_BOUNDS
;
669 case TPT_ERR_INVALIDATE_SHARED_MR
:
670 case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND
:
671 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_OP
;
672 *ecode
= RDMAP_CANT_INV_STAG
;
675 case TPT_ERR_ECC_PSTAG
:
676 case TPT_ERR_INTERNAL_ERR
:
677 *layer_type
= LAYER_RDMAP
|RDMAP_LOCAL_CATA
;
680 case TPT_ERR_OUT_OF_RQE
:
681 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
682 *ecode
= DDPU_INV_MSN_NOBUF
;
684 case TPT_ERR_PBL_ADDR_BOUND
:
685 *layer_type
= LAYER_DDP
|DDP_TAGGED_ERR
;
686 *ecode
= DDPT_BASE_BOUNDS
;
689 *layer_type
= LAYER_MPA
|DDP_LLP
;
690 *ecode
= MPA_CRC_ERR
;
693 *layer_type
= LAYER_MPA
|DDP_LLP
;
694 *ecode
= MPA_MARKER_ERR
;
696 case TPT_ERR_PDU_LEN_ERR
:
697 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
698 *ecode
= DDPU_MSG_TOOBIG
;
700 case TPT_ERR_DDP_VERSION
:
702 *layer_type
= LAYER_DDP
|DDP_TAGGED_ERR
;
703 *ecode
= DDPT_INV_VERS
;
705 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
706 *ecode
= DDPU_INV_VERS
;
709 case TPT_ERR_RDMA_VERSION
:
710 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_OP
;
711 *ecode
= RDMAP_INV_VERS
;
714 *layer_type
= LAYER_RDMAP
|RDMAP_REMOTE_OP
;
715 *ecode
= RDMAP_INV_OPCODE
;
717 case TPT_ERR_DDP_QUEUE_NUM
:
718 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
719 *ecode
= DDPU_INV_QN
;
722 case TPT_ERR_MSN_GAP
:
723 case TPT_ERR_MSN_RANGE
:
724 case TPT_ERR_IRD_OVERFLOW
:
725 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
726 *ecode
= DDPU_INV_MSN_RANGE
;
729 *layer_type
= LAYER_DDP
|DDP_LOCAL_CATA
;
733 *layer_type
= LAYER_DDP
|DDP_UNTAGGED_ERR
;
734 *ecode
= DDPU_INV_MO
;
737 *layer_type
= LAYER_RDMAP
|DDP_LOCAL_CATA
;
743 int iwch_post_zb_read(struct iwch_ep
*ep
)
747 u8 flit_cnt
= sizeof(struct t3_rdma_read_wr
) >> 3;
749 PDBG("%s enter\n", __func__
);
750 skb
= alloc_skb(40, GFP_KERNEL
);
752 printk(KERN_ERR
"%s cannot send zb_read!!\n", __func__
);
755 wqe
= (union t3_wr
*)skb_put(skb
, sizeof(struct t3_rdma_read_wr
));
756 memset(wqe
, 0, sizeof(struct t3_rdma_read_wr
));
757 wqe
->read
.rdmaop
= T3_READ_REQ
;
758 wqe
->read
.reserved
[0] = 0;
759 wqe
->read
.reserved
[1] = 0;
760 wqe
->read
.rem_stag
= cpu_to_be32(1);
761 wqe
->read
.rem_to
= cpu_to_be64(1);
762 wqe
->read
.local_stag
= cpu_to_be32(1);
763 wqe
->read
.local_len
= cpu_to_be32(0);
764 wqe
->read
.local_to
= cpu_to_be64(1);
765 wqe
->send
.wrh
.op_seop_flags
= cpu_to_be32(V_FW_RIWR_OP(T3_WR_READ
));
766 wqe
->send
.wrh
.gen_tid_len
= cpu_to_be32(V_FW_RIWR_TID(ep
->hwtid
)|
767 V_FW_RIWR_LEN(flit_cnt
));
768 skb
->priority
= CPL_PRIORITY_DATA
;
769 return iwch_cxgb3_ofld_send(ep
->com
.qp
->rhp
->rdev
.t3cdev_p
, skb
);
773 * This posts a TERMINATE with layer=RDMA, type=catastrophic.
775 int iwch_post_terminate(struct iwch_qp
*qhp
, struct respQ_msg_t
*rsp_msg
)
778 struct terminate_message
*term
;
781 PDBG("%s %d\n", __func__
, __LINE__
);
782 skb
= alloc_skb(40, GFP_ATOMIC
);
784 printk(KERN_ERR
"%s cannot send TERMINATE!\n", __func__
);
787 wqe
= (union t3_wr
*)skb_put(skb
, 40);
789 wqe
->send
.rdmaop
= T3_TERMINATE
;
791 /* immediate data length */
792 wqe
->send
.plen
= htonl(4);
794 /* immediate data starts here. */
795 term
= (struct terminate_message
*)wqe
->send
.sgl
;
796 build_term_codes(rsp_msg
, &term
->layer_etype
, &term
->ecode
);
797 wqe
->send
.wrh
.op_seop_flags
= cpu_to_be32(V_FW_RIWR_OP(T3_WR_SEND
) |
798 V_FW_RIWR_FLAGS(T3_COMPLETION_FLAG
| T3_NOTIFY_FLAG
));
799 wqe
->send
.wrh
.gen_tid_len
= cpu_to_be32(V_FW_RIWR_TID(qhp
->ep
->hwtid
));
800 skb
->priority
= CPL_PRIORITY_DATA
;
801 return iwch_cxgb3_ofld_send(qhp
->rhp
->rdev
.t3cdev_p
, skb
);
805 * Assumes qhp lock is held.
807 static void __flush_qp(struct iwch_qp
*qhp
, struct iwch_cq
*rchp
,
808 struct iwch_cq
*schp
)
814 PDBG("%s qhp %p rchp %p schp %p\n", __func__
, qhp
, rchp
, schp
);
815 /* take a ref on the qhp since we must release the lock */
816 atomic_inc(&qhp
->refcnt
);
817 spin_unlock(&qhp
->lock
);
819 /* locking hierarchy: cq lock first, then qp lock. */
820 spin_lock(&rchp
->lock
);
821 spin_lock(&qhp
->lock
);
822 cxio_flush_hw_cq(&rchp
->cq
);
823 cxio_count_rcqes(&rchp
->cq
, &qhp
->wq
, &count
);
824 flushed
= cxio_flush_rq(&qhp
->wq
, &rchp
->cq
, count
);
825 spin_unlock(&qhp
->lock
);
826 spin_unlock(&rchp
->lock
);
828 spin_lock(&rchp
->comp_handler_lock
);
829 (*rchp
->ibcq
.comp_handler
)(&rchp
->ibcq
, rchp
->ibcq
.cq_context
);
830 spin_unlock(&rchp
->comp_handler_lock
);
833 /* locking hierarchy: cq lock first, then qp lock. */
834 spin_lock(&schp
->lock
);
835 spin_lock(&qhp
->lock
);
836 cxio_flush_hw_cq(&schp
->cq
);
837 cxio_count_scqes(&schp
->cq
, &qhp
->wq
, &count
);
838 flushed
= cxio_flush_sq(&qhp
->wq
, &schp
->cq
, count
);
839 spin_unlock(&qhp
->lock
);
840 spin_unlock(&schp
->lock
);
842 spin_lock(&schp
->comp_handler_lock
);
843 (*schp
->ibcq
.comp_handler
)(&schp
->ibcq
, schp
->ibcq
.cq_context
);
844 spin_unlock(&schp
->comp_handler_lock
);
848 if (atomic_dec_and_test(&qhp
->refcnt
))
851 spin_lock(&qhp
->lock
);
854 static void flush_qp(struct iwch_qp
*qhp
)
856 struct iwch_cq
*rchp
, *schp
;
858 rchp
= get_chp(qhp
->rhp
, qhp
->attr
.rcq
);
859 schp
= get_chp(qhp
->rhp
, qhp
->attr
.scq
);
861 if (qhp
->ibqp
.uobject
) {
862 cxio_set_wq_in_error(&qhp
->wq
);
863 cxio_set_cq_in_error(&rchp
->cq
);
864 spin_lock(&rchp
->comp_handler_lock
);
865 (*rchp
->ibcq
.comp_handler
)(&rchp
->ibcq
, rchp
->ibcq
.cq_context
);
866 spin_unlock(&rchp
->comp_handler_lock
);
868 cxio_set_cq_in_error(&schp
->cq
);
869 spin_lock(&schp
->comp_handler_lock
);
870 (*schp
->ibcq
.comp_handler
)(&schp
->ibcq
,
871 schp
->ibcq
.cq_context
);
872 spin_unlock(&schp
->comp_handler_lock
);
876 __flush_qp(qhp
, rchp
, schp
);
881 * Return count of RECV WRs posted
883 u16
iwch_rqes_posted(struct iwch_qp
*qhp
)
885 union t3_wr
*wqe
= qhp
->wq
.queue
;
888 while (count
< USHRT_MAX
&& fw_riwrh_opcode((struct fw_riwrh
*)wqe
) == T3_WR_RCV
) {
892 PDBG("%s qhp %p count %u\n", __func__
, qhp
, count
);
896 static int rdma_init(struct iwch_dev
*rhp
, struct iwch_qp
*qhp
,
897 enum iwch_qp_attr_mask mask
,
898 struct iwch_qp_attributes
*attrs
)
900 struct t3_rdma_init_attr init_attr
;
903 init_attr
.tid
= qhp
->ep
->hwtid
;
904 init_attr
.qpid
= qhp
->wq
.qpid
;
905 init_attr
.pdid
= qhp
->attr
.pd
;
906 init_attr
.scqid
= qhp
->attr
.scq
;
907 init_attr
.rcqid
= qhp
->attr
.rcq
;
908 init_attr
.rq_addr
= qhp
->wq
.rq_addr
;
909 init_attr
.rq_size
= 1 << qhp
->wq
.rq_size_log2
;
910 init_attr
.mpaattrs
= uP_RI_MPA_IETF_ENABLE
|
911 qhp
->attr
.mpa_attr
.recv_marker_enabled
|
912 (qhp
->attr
.mpa_attr
.xmit_marker_enabled
<< 1) |
913 (qhp
->attr
.mpa_attr
.crc_enabled
<< 2);
915 init_attr
.qpcaps
= uP_RI_QP_RDMA_READ_ENABLE
|
916 uP_RI_QP_RDMA_WRITE_ENABLE
|
917 uP_RI_QP_BIND_ENABLE
;
918 if (!qhp
->ibqp
.uobject
)
919 init_attr
.qpcaps
|= uP_RI_QP_STAG0_ENABLE
|
920 uP_RI_QP_FAST_REGISTER_ENABLE
;
922 init_attr
.tcp_emss
= qhp
->ep
->emss
;
923 init_attr
.ord
= qhp
->attr
.max_ord
;
924 init_attr
.ird
= qhp
->attr
.max_ird
;
925 init_attr
.qp_dma_addr
= qhp
->wq
.dma_addr
;
926 init_attr
.qp_dma_size
= (1UL << qhp
->wq
.size_log2
);
927 init_attr
.rqe_count
= iwch_rqes_posted(qhp
);
928 init_attr
.flags
= qhp
->attr
.mpa_attr
.initiator
? MPA_INITIATOR
: 0;
929 init_attr
.chan
= qhp
->ep
->l2t
->smt_idx
;
931 init_attr
.rtr_type
= RTR_READ
;
932 if (init_attr
.ord
== 0 && qhp
->attr
.mpa_attr
.initiator
)
934 if (init_attr
.ird
== 0 && !qhp
->attr
.mpa_attr
.initiator
)
937 init_attr
.rtr_type
= 0;
938 init_attr
.irs
= qhp
->ep
->rcv_seq
;
939 PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
940 "flags 0x%x qpcaps 0x%x\n", __func__
,
941 init_attr
.rq_addr
, init_attr
.rq_size
,
942 init_attr
.flags
, init_attr
.qpcaps
);
943 ret
= cxio_rdma_init(&rhp
->rdev
, &init_attr
);
944 PDBG("%s ret %d\n", __func__
, ret
);
948 int iwch_modify_qp(struct iwch_dev
*rhp
, struct iwch_qp
*qhp
,
949 enum iwch_qp_attr_mask mask
,
950 struct iwch_qp_attributes
*attrs
,
954 struct iwch_qp_attributes newattr
= qhp
->attr
;
960 struct iwch_ep
*ep
= NULL
;
962 PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __func__
,
963 qhp
, qhp
->wq
.qpid
, qhp
->ep
, qhp
->attr
.state
,
964 (mask
& IWCH_QP_ATTR_NEXT_STATE
) ? attrs
->next_state
: -1);
966 spin_lock_irqsave(&qhp
->lock
, flag
);
968 /* Process attr changes if in IDLE */
969 if (mask
& IWCH_QP_ATTR_VALID_MODIFY
) {
970 if (qhp
->attr
.state
!= IWCH_QP_STATE_IDLE
) {
974 if (mask
& IWCH_QP_ATTR_ENABLE_RDMA_READ
)
975 newattr
.enable_rdma_read
= attrs
->enable_rdma_read
;
976 if (mask
& IWCH_QP_ATTR_ENABLE_RDMA_WRITE
)
977 newattr
.enable_rdma_write
= attrs
->enable_rdma_write
;
978 if (mask
& IWCH_QP_ATTR_ENABLE_RDMA_BIND
)
979 newattr
.enable_bind
= attrs
->enable_bind
;
980 if (mask
& IWCH_QP_ATTR_MAX_ORD
) {
982 rhp
->attr
.max_rdma_read_qp_depth
) {
986 newattr
.max_ord
= attrs
->max_ord
;
988 if (mask
& IWCH_QP_ATTR_MAX_IRD
) {
990 rhp
->attr
.max_rdma_reads_per_qp
) {
994 newattr
.max_ird
= attrs
->max_ird
;
999 if (!(mask
& IWCH_QP_ATTR_NEXT_STATE
))
1001 if (qhp
->attr
.state
== attrs
->next_state
)
1004 switch (qhp
->attr
.state
) {
1005 case IWCH_QP_STATE_IDLE
:
1006 switch (attrs
->next_state
) {
1007 case IWCH_QP_STATE_RTS
:
1008 if (!(mask
& IWCH_QP_ATTR_LLP_STREAM_HANDLE
)) {
1012 if (!(mask
& IWCH_QP_ATTR_MPA_ATTR
)) {
1016 qhp
->attr
.mpa_attr
= attrs
->mpa_attr
;
1017 qhp
->attr
.llp_stream_handle
= attrs
->llp_stream_handle
;
1018 qhp
->ep
= qhp
->attr
.llp_stream_handle
;
1019 qhp
->attr
.state
= IWCH_QP_STATE_RTS
;
1022 * Ref the endpoint here and deref when we
1023 * disassociate the endpoint from the QP. This
1024 * happens in CLOSING->IDLE transition or *->ERROR
1027 get_ep(&qhp
->ep
->com
);
1028 spin_unlock_irqrestore(&qhp
->lock
, flag
);
1029 ret
= rdma_init(rhp
, qhp
, mask
, attrs
);
1030 spin_lock_irqsave(&qhp
->lock
, flag
);
1034 case IWCH_QP_STATE_ERROR
:
1035 qhp
->attr
.state
= IWCH_QP_STATE_ERROR
;
1043 case IWCH_QP_STATE_RTS
:
1044 switch (attrs
->next_state
) {
1045 case IWCH_QP_STATE_CLOSING
:
1046 BUG_ON(atomic_read(&qhp
->ep
->com
.kref
.refcount
) < 2);
1047 qhp
->attr
.state
= IWCH_QP_STATE_CLOSING
;
1055 case IWCH_QP_STATE_TERMINATE
:
1056 qhp
->attr
.state
= IWCH_QP_STATE_TERMINATE
;
1057 if (qhp
->ibqp
.uobject
)
1058 cxio_set_wq_in_error(&qhp
->wq
);
1062 case IWCH_QP_STATE_ERROR
:
1063 qhp
->attr
.state
= IWCH_QP_STATE_ERROR
;
1077 case IWCH_QP_STATE_CLOSING
:
1082 switch (attrs
->next_state
) {
1083 case IWCH_QP_STATE_IDLE
:
1085 qhp
->attr
.state
= IWCH_QP_STATE_IDLE
;
1086 qhp
->attr
.llp_stream_handle
= NULL
;
1087 put_ep(&qhp
->ep
->com
);
1089 wake_up(&qhp
->wait
);
1091 case IWCH_QP_STATE_ERROR
:
1098 case IWCH_QP_STATE_ERROR
:
1099 if (attrs
->next_state
!= IWCH_QP_STATE_IDLE
) {
1104 if (!Q_EMPTY(qhp
->wq
.sq_rptr
, qhp
->wq
.sq_wptr
) ||
1105 !Q_EMPTY(qhp
->wq
.rq_rptr
, qhp
->wq
.rq_wptr
)) {
1109 qhp
->attr
.state
= IWCH_QP_STATE_IDLE
;
1111 case IWCH_QP_STATE_TERMINATE
:
1119 printk(KERN_ERR
"%s in a bad state %d\n",
1120 __func__
, qhp
->attr
.state
);
1127 PDBG("%s disassociating ep %p qpid 0x%x\n", __func__
, qhp
->ep
,
1130 /* disassociate the LLP connection */
1131 qhp
->attr
.llp_stream_handle
= NULL
;
1134 qhp
->attr
.state
= IWCH_QP_STATE_ERROR
;
1136 wake_up(&qhp
->wait
);
1140 spin_unlock_irqrestore(&qhp
->lock
, flag
);
1143 iwch_post_terminate(qhp
, NULL
);
1146 * If disconnect is 1, then we need to initiate a disconnect
1147 * on the EP. This can be a normal close (RTS->CLOSING) or
1148 * an abnormal close (RTS/CLOSING->ERROR).
1151 iwch_ep_disconnect(ep
, abort
, GFP_KERNEL
);
1156 * If free is 1, then we've disassociated the EP from the QP
1157 * and we need to dereference the EP.
1162 PDBG("%s exit state %d\n", __func__
, qhp
->attr
.state
);