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 <asm/delay.h>
34 #include <linux/mutex.h>
35 #include <linux/netdevice.h>
36 #include <linux/sched.h>
37 #include <linux/spinlock.h>
38 #include <linux/pci.h>
40 #include "cxio_resource.h"
42 #include "cxgb3_offload.h"
45 static LIST_HEAD(rdev_list
);
46 static cxio_hal_ev_callback_func_t cxio_ev_cb
= NULL
;
48 static inline struct cxio_rdev
*cxio_hal_find_rdev_by_name(char *dev_name
)
50 struct cxio_rdev
*rdev
;
52 list_for_each_entry(rdev
, &rdev_list
, entry
)
53 if (!strcmp(rdev
->dev_name
, dev_name
))
58 static inline struct cxio_rdev
*cxio_hal_find_rdev_by_t3cdev(struct t3cdev
61 struct cxio_rdev
*rdev
;
63 list_for_each_entry(rdev
, &rdev_list
, entry
)
64 if (rdev
->t3cdev_p
== tdev
)
69 int cxio_hal_cq_op(struct cxio_rdev
*rdev_p
, struct t3_cq
*cq
,
70 enum t3_cq_opcode op
, u32 credit
)
76 struct rdma_cq_op setup
;
78 setup
.credits
= (op
== CQ_CREDIT_UPDATE
) ? credit
: 0;
80 ret
= rdev_p
->t3cdev_p
->ctl(rdev_p
->t3cdev_p
, RDMA_CQ_OP
, &setup
);
82 if ((ret
< 0) || (op
== CQ_CREDIT_UPDATE
))
86 * If the rearm returned an index other than our current index,
87 * then there might be CQE's in flight (being DMA'd). We must wait
88 * here for them to complete or the consumer can miss a notification.
90 if (Q_PTR2IDX((cq
->rptr
), cq
->size_log2
) != ret
) {
96 * Keep the generation correct by bumping rptr until it
97 * matches the index returned by the rearm - 1.
99 while (Q_PTR2IDX((rptr
+1), cq
->size_log2
) != ret
)
103 * Now rptr is the index for the (last) cqe that was
104 * in-flight at the time the HW rearmed the CQ. We
105 * spin until that CQE is valid.
107 cqe
= cq
->queue
+ Q_PTR2IDX(rptr
, cq
->size_log2
);
108 while (!CQ_VLD_ENTRY(rptr
, cq
->size_log2
, cqe
)) {
112 printk(KERN_ERR
"%s: stalled rnic\n",
121 static inline int cxio_hal_clear_cq_ctx(struct cxio_rdev
*rdev_p
, u32 cqid
)
123 struct rdma_cq_setup setup
;
125 setup
.base_addr
= 0; /* NULL address */
126 setup
.size
= 0; /* disaable the CQ */
128 setup
.credit_thres
= 0;
130 return (rdev_p
->t3cdev_p
->ctl(rdev_p
->t3cdev_p
, RDMA_CQ_SETUP
, &setup
));
133 int cxio_hal_clear_qp_ctx(struct cxio_rdev
*rdev_p
, u32 qpid
)
136 struct t3_modify_qp_wr
*wqe
;
137 struct sk_buff
*skb
= alloc_skb(sizeof(*wqe
), GFP_KERNEL
);
139 PDBG("%s alloc_skb failed\n", __FUNCTION__
);
142 wqe
= (struct t3_modify_qp_wr
*) skb_put(skb
, sizeof(*wqe
));
143 memset(wqe
, 0, sizeof(*wqe
));
144 build_fw_riwrh((struct fw_riwrh
*) wqe
, T3_WR_QP_MOD
, 3, 1, qpid
, 7);
145 wqe
->flags
= cpu_to_be32(MODQP_WRITE_EC
);
146 sge_cmd
= qpid
<< 8 | 3;
147 wqe
->sge_cmd
= cpu_to_be64(sge_cmd
);
148 skb
->priority
= CPL_PRIORITY_CONTROL
;
149 return (cxgb3_ofld_send(rdev_p
->t3cdev_p
, skb
));
152 int cxio_create_cq(struct cxio_rdev
*rdev_p
, struct t3_cq
*cq
)
154 struct rdma_cq_setup setup
;
155 int size
= (1UL << (cq
->size_log2
)) * sizeof(struct t3_cqe
);
157 cq
->cqid
= cxio_hal_get_cqid(rdev_p
->rscp
);
160 cq
->sw_queue
= kzalloc(size
, GFP_KERNEL
);
163 cq
->queue
= dma_alloc_coherent(&(rdev_p
->rnic_info
.pdev
->dev
),
164 (1UL << (cq
->size_log2
)) *
165 sizeof(struct t3_cqe
),
166 &(cq
->dma_addr
), GFP_KERNEL
);
171 pci_unmap_addr_set(cq
, mapping
, cq
->dma_addr
);
172 memset(cq
->queue
, 0, size
);
174 setup
.base_addr
= (u64
) (cq
->dma_addr
);
175 setup
.size
= 1UL << cq
->size_log2
;
176 setup
.credits
= 65535;
177 setup
.credit_thres
= 1;
178 if (rdev_p
->t3cdev_p
->type
== T3B
)
182 return (rdev_p
->t3cdev_p
->ctl(rdev_p
->t3cdev_p
, RDMA_CQ_SETUP
, &setup
));
185 int cxio_resize_cq(struct cxio_rdev
*rdev_p
, struct t3_cq
*cq
)
187 struct rdma_cq_setup setup
;
189 setup
.base_addr
= (u64
) (cq
->dma_addr
);
190 setup
.size
= 1UL << cq
->size_log2
;
191 setup
.credits
= setup
.size
;
192 setup
.credit_thres
= setup
.size
; /* TBD: overflow recovery */
194 return (rdev_p
->t3cdev_p
->ctl(rdev_p
->t3cdev_p
, RDMA_CQ_SETUP
, &setup
));
197 static u32
get_qpid(struct cxio_rdev
*rdev_p
, struct cxio_ucontext
*uctx
)
199 struct cxio_qpid_list
*entry
;
203 mutex_lock(&uctx
->lock
);
204 if (!list_empty(&uctx
->qpids
)) {
205 entry
= list_entry(uctx
->qpids
.next
, struct cxio_qpid_list
,
207 list_del(&entry
->entry
);
211 qpid
= cxio_hal_get_qpid(rdev_p
->rscp
);
214 for (i
= qpid
+1; i
& rdev_p
->qpmask
; i
++) {
215 entry
= kmalloc(sizeof *entry
, GFP_KERNEL
);
219 list_add_tail(&entry
->entry
, &uctx
->qpids
);
223 mutex_unlock(&uctx
->lock
);
224 PDBG("%s qpid 0x%x\n", __FUNCTION__
, qpid
);
228 static void put_qpid(struct cxio_rdev
*rdev_p
, u32 qpid
,
229 struct cxio_ucontext
*uctx
)
231 struct cxio_qpid_list
*entry
;
233 entry
= kmalloc(sizeof *entry
, GFP_KERNEL
);
236 PDBG("%s qpid 0x%x\n", __FUNCTION__
, qpid
);
238 mutex_lock(&uctx
->lock
);
239 list_add_tail(&entry
->entry
, &uctx
->qpids
);
240 mutex_unlock(&uctx
->lock
);
243 void cxio_release_ucontext(struct cxio_rdev
*rdev_p
, struct cxio_ucontext
*uctx
)
245 struct list_head
*pos
, *nxt
;
246 struct cxio_qpid_list
*entry
;
248 mutex_lock(&uctx
->lock
);
249 list_for_each_safe(pos
, nxt
, &uctx
->qpids
) {
250 entry
= list_entry(pos
, struct cxio_qpid_list
, entry
);
251 list_del_init(&entry
->entry
);
252 if (!(entry
->qpid
& rdev_p
->qpmask
))
253 cxio_hal_put_qpid(rdev_p
->rscp
, entry
->qpid
);
256 mutex_unlock(&uctx
->lock
);
259 void cxio_init_ucontext(struct cxio_rdev
*rdev_p
, struct cxio_ucontext
*uctx
)
261 INIT_LIST_HEAD(&uctx
->qpids
);
262 mutex_init(&uctx
->lock
);
265 int cxio_create_qp(struct cxio_rdev
*rdev_p
, u32 kernel_domain
,
266 struct t3_wq
*wq
, struct cxio_ucontext
*uctx
)
268 int depth
= 1UL << wq
->size_log2
;
269 int rqsize
= 1UL << wq
->rq_size_log2
;
271 wq
->qpid
= get_qpid(rdev_p
, uctx
);
275 wq
->rq
= kzalloc(depth
* sizeof(u64
), GFP_KERNEL
);
279 wq
->rq_addr
= cxio_hal_rqtpool_alloc(rdev_p
, rqsize
);
283 wq
->sq
= kzalloc(depth
* sizeof(struct t3_swsq
), GFP_KERNEL
);
287 wq
->queue
= dma_alloc_coherent(&(rdev_p
->rnic_info
.pdev
->dev
),
288 depth
* sizeof(union t3_wr
),
289 &(wq
->dma_addr
), GFP_KERNEL
);
293 memset(wq
->queue
, 0, depth
* sizeof(union t3_wr
));
294 pci_unmap_addr_set(wq
, mapping
, wq
->dma_addr
);
295 wq
->doorbell
= (void __iomem
*)rdev_p
->rnic_info
.kdb_addr
;
297 wq
->udb
= (u64
)rdev_p
->rnic_info
.udbell_physbase
+
298 (wq
->qpid
<< rdev_p
->qpshift
);
299 PDBG("%s qpid 0x%x doorbell 0x%p udb 0x%llx\n", __FUNCTION__
,
300 wq
->qpid
, wq
->doorbell
, (unsigned long long) wq
->udb
);
305 cxio_hal_rqtpool_free(rdev_p
, wq
->rq_addr
, rqsize
);
309 put_qpid(rdev_p
, wq
->qpid
, uctx
);
313 int cxio_destroy_cq(struct cxio_rdev
*rdev_p
, struct t3_cq
*cq
)
316 err
= cxio_hal_clear_cq_ctx(rdev_p
, cq
->cqid
);
318 dma_free_coherent(&(rdev_p
->rnic_info
.pdev
->dev
),
319 (1UL << (cq
->size_log2
))
320 * sizeof(struct t3_cqe
), cq
->queue
,
321 pci_unmap_addr(cq
, mapping
));
322 cxio_hal_put_cqid(rdev_p
->rscp
, cq
->cqid
);
326 int cxio_destroy_qp(struct cxio_rdev
*rdev_p
, struct t3_wq
*wq
,
327 struct cxio_ucontext
*uctx
)
329 dma_free_coherent(&(rdev_p
->rnic_info
.pdev
->dev
),
330 (1UL << (wq
->size_log2
))
331 * sizeof(union t3_wr
), wq
->queue
,
332 pci_unmap_addr(wq
, mapping
));
334 cxio_hal_rqtpool_free(rdev_p
, wq
->rq_addr
, (1UL << wq
->rq_size_log2
));
336 put_qpid(rdev_p
, wq
->qpid
, uctx
);
340 static void insert_recv_cqe(struct t3_wq
*wq
, struct t3_cq
*cq
)
344 PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __FUNCTION__
,
345 wq
, cq
, cq
->sw_rptr
, cq
->sw_wptr
);
346 memset(&cqe
, 0, sizeof(cqe
));
347 cqe
.header
= cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH
) |
348 V_CQE_OPCODE(T3_SEND
) |
351 V_CQE_QPID(wq
->qpid
) |
352 V_CQE_GENBIT(Q_GENBIT(cq
->sw_wptr
,
354 *(cq
->sw_queue
+ Q_PTR2IDX(cq
->sw_wptr
, cq
->size_log2
)) = cqe
;
358 void cxio_flush_rq(struct t3_wq
*wq
, struct t3_cq
*cq
, int count
)
362 PDBG("%s wq %p cq %p\n", __FUNCTION__
, wq
, cq
);
365 PDBG("%s rq_rptr %u rq_wptr %u skip count %u\n", __FUNCTION__
,
366 wq
->rq_rptr
, wq
->rq_wptr
, count
);
367 ptr
= wq
->rq_rptr
+ count
;
368 while (ptr
++ != wq
->rq_wptr
)
369 insert_recv_cqe(wq
, cq
);
372 static void insert_sq_cqe(struct t3_wq
*wq
, struct t3_cq
*cq
,
377 PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __FUNCTION__
,
378 wq
, cq
, cq
->sw_rptr
, cq
->sw_wptr
);
379 memset(&cqe
, 0, sizeof(cqe
));
380 cqe
.header
= cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH
) |
381 V_CQE_OPCODE(sqp
->opcode
) |
384 V_CQE_QPID(wq
->qpid
) |
385 V_CQE_GENBIT(Q_GENBIT(cq
->sw_wptr
,
387 cqe
.u
.scqe
.wrid_hi
= sqp
->sq_wptr
;
389 *(cq
->sw_queue
+ Q_PTR2IDX(cq
->sw_wptr
, cq
->size_log2
)) = cqe
;
393 void cxio_flush_sq(struct t3_wq
*wq
, struct t3_cq
*cq
, int count
)
396 struct t3_swsq
*sqp
= wq
->sq
+ Q_PTR2IDX(wq
->sq_rptr
, wq
->sq_size_log2
);
398 ptr
= wq
->sq_rptr
+ count
;
400 while (ptr
!= wq
->sq_wptr
) {
401 insert_sq_cqe(wq
, cq
, sqp
);
408 * Move all CQEs from the HWCQ into the SWCQ.
410 void cxio_flush_hw_cq(struct t3_cq
*cq
)
412 struct t3_cqe
*cqe
, *swcqe
;
414 PDBG("%s cq %p cqid 0x%x\n", __FUNCTION__
, cq
, cq
->cqid
);
415 cqe
= cxio_next_hw_cqe(cq
);
417 PDBG("%s flushing hwcq rptr 0x%x to swcq wptr 0x%x\n",
418 __FUNCTION__
, cq
->rptr
, cq
->sw_wptr
);
419 swcqe
= cq
->sw_queue
+ Q_PTR2IDX(cq
->sw_wptr
, cq
->size_log2
);
421 swcqe
->header
|= cpu_to_be32(V_CQE_SWCQE(1));
424 cqe
= cxio_next_hw_cqe(cq
);
428 static inline int cqe_completes_wr(struct t3_cqe
*cqe
, struct t3_wq
*wq
)
430 if (CQE_OPCODE(*cqe
) == T3_TERMINATE
)
433 if ((CQE_OPCODE(*cqe
) == T3_RDMA_WRITE
) && RQ_TYPE(*cqe
))
436 if ((CQE_OPCODE(*cqe
) == T3_READ_RESP
) && SQ_TYPE(*cqe
))
439 if ((CQE_OPCODE(*cqe
) == T3_SEND
) && RQ_TYPE(*cqe
) &&
440 Q_EMPTY(wq
->rq_rptr
, wq
->rq_wptr
))
446 void cxio_count_scqes(struct t3_cq
*cq
, struct t3_wq
*wq
, int *count
)
453 while (!Q_EMPTY(ptr
, cq
->sw_wptr
)) {
454 cqe
= cq
->sw_queue
+ (Q_PTR2IDX(ptr
, cq
->size_log2
));
455 if ((SQ_TYPE(*cqe
) || (CQE_OPCODE(*cqe
) == T3_READ_RESP
)) &&
456 (CQE_QPID(*cqe
) == wq
->qpid
))
460 PDBG("%s cq %p count %d\n", __FUNCTION__
, cq
, *count
);
463 void cxio_count_rcqes(struct t3_cq
*cq
, struct t3_wq
*wq
, int *count
)
469 PDBG("%s count zero %d\n", __FUNCTION__
, *count
);
471 while (!Q_EMPTY(ptr
, cq
->sw_wptr
)) {
472 cqe
= cq
->sw_queue
+ (Q_PTR2IDX(ptr
, cq
->size_log2
));
473 if (RQ_TYPE(*cqe
) && (CQE_OPCODE(*cqe
) != T3_READ_RESP
) &&
474 (CQE_QPID(*cqe
) == wq
->qpid
) && cqe_completes_wr(cqe
, wq
))
478 PDBG("%s cq %p count %d\n", __FUNCTION__
, cq
, *count
);
481 static int cxio_hal_init_ctrl_cq(struct cxio_rdev
*rdev_p
)
483 struct rdma_cq_setup setup
;
485 setup
.base_addr
= 0; /* NULL address */
486 setup
.size
= 1; /* enable the CQ */
489 /* force SGE to redirect to RspQ and interrupt */
490 setup
.credit_thres
= 0;
492 return (rdev_p
->t3cdev_p
->ctl(rdev_p
->t3cdev_p
, RDMA_CQ_SETUP
, &setup
));
495 static int cxio_hal_init_ctrl_qp(struct cxio_rdev
*rdev_p
)
498 u64 sge_cmd
, ctx0
, ctx1
;
500 struct t3_modify_qp_wr
*wqe
;
501 struct sk_buff
*skb
= alloc_skb(sizeof(*wqe
), GFP_KERNEL
);
505 PDBG("%s alloc_skb failed\n", __FUNCTION__
);
508 err
= cxio_hal_init_ctrl_cq(rdev_p
);
510 PDBG("%s err %d initializing ctrl_cq\n", __FUNCTION__
, err
);
513 rdev_p
->ctrl_qp
.workq
= dma_alloc_coherent(
514 &(rdev_p
->rnic_info
.pdev
->dev
),
515 (1 << T3_CTRL_QP_SIZE_LOG2
) *
517 &(rdev_p
->ctrl_qp
.dma_addr
),
519 if (!rdev_p
->ctrl_qp
.workq
) {
520 PDBG("%s dma_alloc_coherent failed\n", __FUNCTION__
);
523 pci_unmap_addr_set(&rdev_p
->ctrl_qp
, mapping
,
524 rdev_p
->ctrl_qp
.dma_addr
);
525 rdev_p
->ctrl_qp
.doorbell
= (void __iomem
*)rdev_p
->rnic_info
.kdb_addr
;
526 memset(rdev_p
->ctrl_qp
.workq
, 0,
527 (1 << T3_CTRL_QP_SIZE_LOG2
) * sizeof(union t3_wr
));
529 mutex_init(&rdev_p
->ctrl_qp
.lock
);
530 init_waitqueue_head(&rdev_p
->ctrl_qp
.waitq
);
532 /* update HW Ctrl QP context */
533 base_addr
= rdev_p
->ctrl_qp
.dma_addr
;
535 ctx0
= (V_EC_SIZE((1 << T3_CTRL_QP_SIZE_LOG2
)) |
536 V_EC_BASE_LO((u32
) base_addr
& 0xffff));
538 ctx0
|= V_EC_CREDITS(FW_WR_NUM
);
540 ctx1
= (u32
) base_addr
;
542 ctx1
|= ((u64
) (V_EC_BASE_HI((u32
) base_addr
& 0xf) | V_EC_RESPQ(0) |
543 V_EC_TYPE(0) | V_EC_GEN(1) |
544 V_EC_UP_TOKEN(T3_CTL_QP_TID
) | F_EC_VALID
)) << 32;
545 wqe
= (struct t3_modify_qp_wr
*) skb_put(skb
, sizeof(*wqe
));
546 memset(wqe
, 0, sizeof(*wqe
));
547 build_fw_riwrh((struct fw_riwrh
*) wqe
, T3_WR_QP_MOD
, 0, 1,
549 wqe
->flags
= cpu_to_be32(MODQP_WRITE_EC
);
550 sge_cmd
= (3ULL << 56) | FW_RI_SGEEC_START
<< 8 | 3;
551 wqe
->sge_cmd
= cpu_to_be64(sge_cmd
);
552 wqe
->ctx1
= cpu_to_be64(ctx1
);
553 wqe
->ctx0
= cpu_to_be64(ctx0
);
554 PDBG("CtrlQP dma_addr 0x%llx workq %p size %d\n",
555 (unsigned long long) rdev_p
->ctrl_qp
.dma_addr
,
556 rdev_p
->ctrl_qp
.workq
, 1 << T3_CTRL_QP_SIZE_LOG2
);
557 skb
->priority
= CPL_PRIORITY_CONTROL
;
558 return (cxgb3_ofld_send(rdev_p
->t3cdev_p
, skb
));
561 static int cxio_hal_destroy_ctrl_qp(struct cxio_rdev
*rdev_p
)
563 dma_free_coherent(&(rdev_p
->rnic_info
.pdev
->dev
),
564 (1UL << T3_CTRL_QP_SIZE_LOG2
)
565 * sizeof(union t3_wr
), rdev_p
->ctrl_qp
.workq
,
566 pci_unmap_addr(&rdev_p
->ctrl_qp
, mapping
));
567 return cxio_hal_clear_qp_ctx(rdev_p
, T3_CTRL_QP_ID
);
570 /* write len bytes of data into addr (32B aligned address)
571 * If data is NULL, clear len byte of memory to zero.
572 * caller aquires the ctrl_qp lock before the call
574 static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev
*rdev_p
, u32 addr
,
575 u32 len
, void *data
, int completion
)
577 u32 i
, nr_wqe
, copy_len
;
579 u8 wr_len
, utx_len
; /* lenght in 8 byte flit */
580 enum t3_wr_flags flag
;
584 nr_wqe
= len
% 96 ? len
/ 96 + 1 : len
/ 96; /* 96B max per WQE */
585 PDBG("%s wptr 0x%x rptr 0x%x len %d, nr_wqe %d data %p addr 0x%0x\n",
586 __FUNCTION__
, rdev_p
->ctrl_qp
.wptr
, rdev_p
->ctrl_qp
.rptr
, len
,
588 utx_len
= 3; /* in 32B unit */
589 for (i
= 0; i
< nr_wqe
; i
++) {
590 if (Q_FULL(rdev_p
->ctrl_qp
.rptr
, rdev_p
->ctrl_qp
.wptr
,
591 T3_CTRL_QP_SIZE_LOG2
)) {
592 PDBG("%s ctrl_qp full wtpr 0x%0x rptr 0x%0x, "
593 "wait for more space i %d\n", __FUNCTION__
,
594 rdev_p
->ctrl_qp
.wptr
, rdev_p
->ctrl_qp
.rptr
, i
);
595 if (wait_event_interruptible(rdev_p
->ctrl_qp
.waitq
,
596 !Q_FULL(rdev_p
->ctrl_qp
.rptr
,
597 rdev_p
->ctrl_qp
.wptr
,
598 T3_CTRL_QP_SIZE_LOG2
))) {
599 PDBG("%s ctrl_qp workq interrupted\n",
603 PDBG("%s ctrl_qp wakeup, continue posting work request "
604 "i %d\n", __FUNCTION__
, i
);
606 wqe
= (__be64
*)(rdev_p
->ctrl_qp
.workq
+ (rdev_p
->ctrl_qp
.wptr
%
607 (1 << T3_CTRL_QP_SIZE_LOG2
)));
609 if (i
== (nr_wqe
- 1)) {
611 flag
= completion
? T3_COMPLETION_FLAG
: 0;
613 utx_len
= len
/ 32 + 1;
619 * Force a CQE to return the credit to the workq in case
620 * we posted more than half the max QP size of WRs
623 (i
% (((1 << T3_CTRL_QP_SIZE_LOG2
)) >> 1) == 0)) {
624 flag
= T3_COMPLETION_FLAG
;
625 PDBG("%s force completion at i %d\n", __FUNCTION__
, i
);
628 /* build the utx mem command */
629 wqe
+= (sizeof(struct t3_bypass_wr
) >> 3);
630 utx_cmd
= (T3_UTX_MEM_WRITE
<< 28) | (addr
+ i
* 3);
632 utx_cmd
|= (utx_len
<< 28) | ((utx_len
<< 2) + 1);
633 *wqe
= cpu_to_be64(utx_cmd
);
635 copy_data
= (u8
*) data
+ i
* 96;
636 copy_len
= len
> 96 ? 96 : len
;
638 /* clear memory content if data is NULL */
640 memcpy(wqe
, copy_data
, copy_len
);
642 memset(wqe
, 0, copy_len
);
644 memset(((u8
*) wqe
) + copy_len
, 0,
645 32 - (copy_len
% 32));
646 wr_len
= ((sizeof(struct t3_bypass_wr
)) >> 3) + 1 +
648 wqe
= (__be64
*)(rdev_p
->ctrl_qp
.workq
+ (rdev_p
->ctrl_qp
.wptr
%
649 (1 << T3_CTRL_QP_SIZE_LOG2
)));
651 /* wptr in the WRID[31:0] */
652 ((union t3_wrid
*)(wqe
+1))->id0
.low
= rdev_p
->ctrl_qp
.wptr
;
655 * This must be the last write with a memory barrier
658 build_fw_riwrh((struct fw_riwrh
*) wqe
, T3_WR_BP
, flag
,
659 Q_GENBIT(rdev_p
->ctrl_qp
.wptr
,
660 T3_CTRL_QP_SIZE_LOG2
), T3_CTRL_QP_ID
,
662 if (flag
== T3_COMPLETION_FLAG
)
663 ring_doorbell(rdev_p
->ctrl_qp
.doorbell
, T3_CTRL_QP_ID
);
665 rdev_p
->ctrl_qp
.wptr
++;
670 /* IN: stag key, pdid, perm, zbva, to, len, page_size, pbl, and pbl_size
671 * OUT: stag index, actual pbl_size, pbl_addr allocated.
672 * TBD: shared memory region support
674 static int __cxio_tpt_op(struct cxio_rdev
*rdev_p
, u32 reset_tpt_entry
,
675 u32
*stag
, u8 stag_state
, u32 pdid
,
676 enum tpt_mem_type type
, enum tpt_mem_perm perm
,
677 u32 zbva
, u64 to
, u32 len
, u8 page_size
, __be64
*pbl
,
678 u32
*pbl_size
, u32
*pbl_addr
)
681 struct tpt_entry tpt
;
684 int rereg
= (*stag
!= T3_STAG_UNSET
);
686 stag_state
= stag_state
> 0;
687 stag_idx
= (*stag
) >> 8;
689 if ((!reset_tpt_entry
) && !(*stag
!= T3_STAG_UNSET
)) {
690 stag_idx
= cxio_hal_get_stag(rdev_p
->rscp
);
693 *stag
= (stag_idx
<< 8) | ((*stag
) & 0xFF);
695 PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
696 __FUNCTION__
, stag_state
, type
, pdid
, stag_idx
);
699 cxio_hal_pblpool_free(rdev_p
, *pbl_addr
, *pbl_size
<< 3);
701 *pbl_addr
= cxio_hal_pblpool_alloc(rdev_p
, *pbl_size
<< 3);
707 mutex_lock(&rdev_p
->ctrl_qp
.lock
);
709 /* write PBL first if any - update pbl only if pbl list exist */
712 PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
713 __FUNCTION__
, *pbl_addr
, rdev_p
->rnic_info
.pbl_base
,
715 err
= cxio_hal_ctrl_qp_write_mem(rdev_p
,
717 (*pbl_size
<< 3), pbl
, 0);
722 /* write TPT entry */
724 memset(&tpt
, 0, sizeof(tpt
));
726 tpt
.valid_stag_pdid
= cpu_to_be32(F_TPT_VALID
|
727 V_TPT_STAG_KEY((*stag
) & M_TPT_STAG_KEY
) |
728 V_TPT_STAG_STATE(stag_state
) |
729 V_TPT_STAG_TYPE(type
) | V_TPT_PDID(pdid
));
730 BUG_ON(page_size
>= 28);
731 tpt
.flags_pagesize_qpid
= cpu_to_be32(V_TPT_PERM(perm
) |
732 F_TPT_MW_BIND_ENABLE
|
733 V_TPT_ADDR_TYPE((zbva
? TPT_ZBTO
: TPT_VATO
)) |
734 V_TPT_PAGE_SIZE(page_size
));
735 tpt
.rsvd_pbl_addr
= reset_tpt_entry
? 0 :
736 cpu_to_be32(V_TPT_PBL_ADDR(PBL_OFF(rdev_p
, *pbl_addr
)>>3));
737 tpt
.len
= cpu_to_be32(len
);
738 tpt
.va_hi
= cpu_to_be32((u32
) (to
>> 32));
739 tpt
.va_low_or_fbo
= cpu_to_be32((u32
) (to
& 0xFFFFFFFFULL
));
740 tpt
.rsvd_bind_cnt_or_pstag
= 0;
741 tpt
.rsvd_pbl_size
= reset_tpt_entry
? 0 :
742 cpu_to_be32(V_TPT_PBL_SIZE((*pbl_size
) >> 2));
744 err
= cxio_hal_ctrl_qp_write_mem(rdev_p
,
746 (rdev_p
->rnic_info
.tpt_base
>> 5),
747 sizeof(tpt
), &tpt
, 1);
749 /* release the stag index to free pool */
751 cxio_hal_put_stag(rdev_p
->rscp
, stag_idx
);
753 wptr
= rdev_p
->ctrl_qp
.wptr
;
754 mutex_unlock(&rdev_p
->ctrl_qp
.lock
);
756 if (wait_event_interruptible(rdev_p
->ctrl_qp
.waitq
,
757 SEQ32_GE(rdev_p
->ctrl_qp
.rptr
,
763 /* IN : stag key, pdid, pbl_size
764 * Out: stag index, actaul pbl_size, and pbl_addr allocated.
766 int cxio_allocate_stag(struct cxio_rdev
*rdev_p
, u32
* stag
, u32 pdid
,
767 enum tpt_mem_perm perm
, u32
* pbl_size
, u32
* pbl_addr
)
769 *stag
= T3_STAG_UNSET
;
770 return (__cxio_tpt_op(rdev_p
, 0, stag
, 0, pdid
, TPT_NON_SHARED_MR
,
771 perm
, 0, 0ULL, 0, 0, NULL
, pbl_size
, pbl_addr
));
774 int cxio_register_phys_mem(struct cxio_rdev
*rdev_p
, u32
*stag
, u32 pdid
,
775 enum tpt_mem_perm perm
, u32 zbva
, u64 to
, u32 len
,
776 u8 page_size
, __be64
*pbl
, u32
*pbl_size
,
779 *stag
= T3_STAG_UNSET
;
780 return __cxio_tpt_op(rdev_p
, 0, stag
, 1, pdid
, TPT_NON_SHARED_MR
, perm
,
781 zbva
, to
, len
, page_size
, pbl
, pbl_size
, pbl_addr
);
784 int cxio_reregister_phys_mem(struct cxio_rdev
*rdev_p
, u32
*stag
, u32 pdid
,
785 enum tpt_mem_perm perm
, u32 zbva
, u64 to
, u32 len
,
786 u8 page_size
, __be64
*pbl
, u32
*pbl_size
,
789 return __cxio_tpt_op(rdev_p
, 0, stag
, 1, pdid
, TPT_NON_SHARED_MR
, perm
,
790 zbva
, to
, len
, page_size
, pbl
, pbl_size
, pbl_addr
);
793 int cxio_dereg_mem(struct cxio_rdev
*rdev_p
, u32 stag
, u32 pbl_size
,
796 return __cxio_tpt_op(rdev_p
, 1, &stag
, 0, 0, 0, 0, 0, 0ULL, 0, 0, NULL
,
797 &pbl_size
, &pbl_addr
);
800 int cxio_allocate_window(struct cxio_rdev
*rdev_p
, u32
* stag
, u32 pdid
)
803 *stag
= T3_STAG_UNSET
;
804 return __cxio_tpt_op(rdev_p
, 0, stag
, 0, pdid
, TPT_MW
, 0, 0, 0ULL, 0, 0,
805 NULL
, &pbl_size
, NULL
);
808 int cxio_deallocate_window(struct cxio_rdev
*rdev_p
, u32 stag
)
810 return __cxio_tpt_op(rdev_p
, 1, &stag
, 0, 0, 0, 0, 0, 0ULL, 0, 0, NULL
,
814 int cxio_rdma_init(struct cxio_rdev
*rdev_p
, struct t3_rdma_init_attr
*attr
)
816 struct t3_rdma_init_wr
*wqe
;
817 struct sk_buff
*skb
= alloc_skb(sizeof(*wqe
), GFP_ATOMIC
);
820 PDBG("%s rdev_p %p\n", __FUNCTION__
, rdev_p
);
821 wqe
= (struct t3_rdma_init_wr
*) __skb_put(skb
, sizeof(*wqe
));
822 wqe
->wrh
.op_seop_flags
= cpu_to_be32(V_FW_RIWR_OP(T3_WR_INIT
));
823 wqe
->wrh
.gen_tid_len
= cpu_to_be32(V_FW_RIWR_TID(attr
->tid
) |
824 V_FW_RIWR_LEN(sizeof(*wqe
) >> 3));
826 wqe
->qpid
= cpu_to_be32(attr
->qpid
);
827 wqe
->pdid
= cpu_to_be32(attr
->pdid
);
828 wqe
->scqid
= cpu_to_be32(attr
->scqid
);
829 wqe
->rcqid
= cpu_to_be32(attr
->rcqid
);
830 wqe
->rq_addr
= cpu_to_be32(attr
->rq_addr
- rdev_p
->rnic_info
.rqt_base
);
831 wqe
->rq_size
= cpu_to_be32(attr
->rq_size
);
832 wqe
->mpaattrs
= attr
->mpaattrs
;
833 wqe
->qpcaps
= attr
->qpcaps
;
834 wqe
->ulpdu_size
= cpu_to_be16(attr
->tcp_emss
);
835 wqe
->flags
= cpu_to_be32(attr
->flags
);
836 wqe
->ord
= cpu_to_be32(attr
->ord
);
837 wqe
->ird
= cpu_to_be32(attr
->ird
);
838 wqe
->qp_dma_addr
= cpu_to_be64(attr
->qp_dma_addr
);
839 wqe
->qp_dma_size
= cpu_to_be32(attr
->qp_dma_size
);
841 skb
->priority
= 0; /* 0=>ToeQ; 1=>CtrlQ */
842 return (cxgb3_ofld_send(rdev_p
->t3cdev_p
, skb
));
845 void cxio_register_ev_cb(cxio_hal_ev_callback_func_t ev_cb
)
850 void cxio_unregister_ev_cb(cxio_hal_ev_callback_func_t ev_cb
)
855 static int cxio_hal_ev_handler(struct t3cdev
*t3cdev_p
, struct sk_buff
*skb
)
858 struct cxio_rdev
*rdev_p
= NULL
;
859 struct respQ_msg_t
*rsp_msg
= (struct respQ_msg_t
*) skb
->data
;
860 PDBG("%d: %s cq_id 0x%x cq_ptr 0x%x genbit %0x overflow %0x an %0x"
861 " se %0x notify %0x cqbranch %0x creditth %0x\n",
862 cnt
, __FUNCTION__
, RSPQ_CQID(rsp_msg
), RSPQ_CQPTR(rsp_msg
),
863 RSPQ_GENBIT(rsp_msg
), RSPQ_OVERFLOW(rsp_msg
), RSPQ_AN(rsp_msg
),
864 RSPQ_SE(rsp_msg
), RSPQ_NOTIFY(rsp_msg
), RSPQ_CQBRANCH(rsp_msg
),
865 RSPQ_CREDIT_THRESH(rsp_msg
));
866 PDBG("CQE: QPID 0x%0x genbit %0x type 0x%0x status 0x%0x opcode %d "
867 "len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
868 CQE_QPID(rsp_msg
->cqe
), CQE_GENBIT(rsp_msg
->cqe
),
869 CQE_TYPE(rsp_msg
->cqe
), CQE_STATUS(rsp_msg
->cqe
),
870 CQE_OPCODE(rsp_msg
->cqe
), CQE_LEN(rsp_msg
->cqe
),
871 CQE_WRID_HI(rsp_msg
->cqe
), CQE_WRID_LOW(rsp_msg
->cqe
));
872 rdev_p
= (struct cxio_rdev
*)t3cdev_p
->ulp
;
874 PDBG("%s called by t3cdev %p with null ulp\n", __FUNCTION__
,
878 if (CQE_QPID(rsp_msg
->cqe
) == T3_CTRL_QP_ID
) {
879 rdev_p
->ctrl_qp
.rptr
= CQE_WRID_LOW(rsp_msg
->cqe
) + 1;
880 wake_up_interruptible(&rdev_p
->ctrl_qp
.waitq
);
881 dev_kfree_skb_irq(skb
);
882 } else if (CQE_QPID(rsp_msg
->cqe
) == 0xfff8)
883 dev_kfree_skb_irq(skb
);
885 (*cxio_ev_cb
) (rdev_p
, skb
);
887 dev_kfree_skb_irq(skb
);
892 /* Caller takes care of locking if needed */
893 int cxio_rdev_open(struct cxio_rdev
*rdev_p
)
895 struct net_device
*netdev_p
= NULL
;
897 if (strlen(rdev_p
->dev_name
)) {
898 if (cxio_hal_find_rdev_by_name(rdev_p
->dev_name
)) {
901 netdev_p
= dev_get_by_name(rdev_p
->dev_name
);
906 } else if (rdev_p
->t3cdev_p
) {
907 if (cxio_hal_find_rdev_by_t3cdev(rdev_p
->t3cdev_p
)) {
910 netdev_p
= rdev_p
->t3cdev_p
->lldev
;
911 strncpy(rdev_p
->dev_name
, rdev_p
->t3cdev_p
->name
,
912 T3_MAX_DEV_NAME_LEN
);
914 PDBG("%s t3cdev_p or dev_name must be set\n", __FUNCTION__
);
918 list_add_tail(&rdev_p
->entry
, &rdev_list
);
920 PDBG("%s opening rnic dev %s\n", __FUNCTION__
, rdev_p
->dev_name
);
921 memset(&rdev_p
->ctrl_qp
, 0, sizeof(rdev_p
->ctrl_qp
));
922 if (!rdev_p
->t3cdev_p
)
923 rdev_p
->t3cdev_p
= T3CDEV(netdev_p
);
924 rdev_p
->t3cdev_p
->ulp
= (void *) rdev_p
;
925 err
= rdev_p
->t3cdev_p
->ctl(rdev_p
->t3cdev_p
, RDMA_GET_PARAMS
,
926 &(rdev_p
->rnic_info
));
928 printk(KERN_ERR
"%s t3cdev_p(%p)->ctl returned error %d.\n",
929 __FUNCTION__
, rdev_p
->t3cdev_p
, err
);
932 err
= rdev_p
->t3cdev_p
->ctl(rdev_p
->t3cdev_p
, GET_PORTS
,
933 &(rdev_p
->port_info
));
935 printk(KERN_ERR
"%s t3cdev_p(%p)->ctl returned error %d.\n",
936 __FUNCTION__
, rdev_p
->t3cdev_p
, err
);
941 * qpshift is the number of bits to shift the qpid left in order
942 * to get the correct address of the doorbell for that qp.
944 cxio_init_ucontext(rdev_p
, &rdev_p
->uctx
);
945 rdev_p
->qpshift
= PAGE_SHIFT
-
947 ilog2(rdev_p
->rnic_info
.udbell_len
>>
949 rdev_p
->qpnr
= rdev_p
->rnic_info
.udbell_len
>> PAGE_SHIFT
;
950 rdev_p
->qpmask
= (65536 >> ilog2(rdev_p
->qpnr
)) - 1;
951 PDBG("%s rnic %s info: tpt_base 0x%0x tpt_top 0x%0x num stags %d "
952 "pbl_base 0x%0x pbl_top 0x%0x rqt_base 0x%0x, rqt_top 0x%0x\n",
953 __FUNCTION__
, rdev_p
->dev_name
, rdev_p
->rnic_info
.tpt_base
,
954 rdev_p
->rnic_info
.tpt_top
, cxio_num_stags(rdev_p
),
955 rdev_p
->rnic_info
.pbl_base
,
956 rdev_p
->rnic_info
.pbl_top
, rdev_p
->rnic_info
.rqt_base
,
957 rdev_p
->rnic_info
.rqt_top
);
958 PDBG("udbell_len 0x%0x udbell_physbase 0x%lx kdb_addr %p qpshift %lu "
959 "qpnr %d qpmask 0x%x\n",
960 rdev_p
->rnic_info
.udbell_len
,
961 rdev_p
->rnic_info
.udbell_physbase
, rdev_p
->rnic_info
.kdb_addr
,
962 rdev_p
->qpshift
, rdev_p
->qpnr
, rdev_p
->qpmask
);
964 err
= cxio_hal_init_ctrl_qp(rdev_p
);
966 printk(KERN_ERR
"%s error %d initializing ctrl_qp.\n",
970 err
= cxio_hal_init_resource(rdev_p
, cxio_num_stags(rdev_p
), 0,
971 0, T3_MAX_NUM_QP
, T3_MAX_NUM_CQ
,
974 printk(KERN_ERR
"%s error %d initializing hal resources.\n",
978 err
= cxio_hal_pblpool_create(rdev_p
);
980 printk(KERN_ERR
"%s error %d initializing pbl mem pool.\n",
984 err
= cxio_hal_rqtpool_create(rdev_p
);
986 printk(KERN_ERR
"%s error %d initializing rqt mem pool.\n",
992 cxio_hal_pblpool_destroy(rdev_p
);
994 cxio_hal_destroy_resource(rdev_p
->rscp
);
996 cxio_hal_destroy_ctrl_qp(rdev_p
);
998 list_del(&rdev_p
->entry
);
1002 void cxio_rdev_close(struct cxio_rdev
*rdev_p
)
1005 cxio_hal_pblpool_destroy(rdev_p
);
1006 cxio_hal_rqtpool_destroy(rdev_p
);
1007 list_del(&rdev_p
->entry
);
1008 rdev_p
->t3cdev_p
->ulp
= NULL
;
1009 cxio_hal_destroy_ctrl_qp(rdev_p
);
1010 cxio_hal_destroy_resource(rdev_p
->rscp
);
1014 int __init
cxio_hal_init(void)
1016 if (cxio_hal_init_rhdl_resource(T3_MAX_NUM_RI
))
1018 t3_register_cpl_handler(CPL_ASYNC_NOTIF
, cxio_hal_ev_handler
);
1022 void __exit
cxio_hal_exit(void)
1024 struct cxio_rdev
*rdev
, *tmp
;
1026 t3_register_cpl_handler(CPL_ASYNC_NOTIF
, NULL
);
1027 list_for_each_entry_safe(rdev
, tmp
, &rdev_list
, entry
)
1028 cxio_rdev_close(rdev
);
1029 cxio_hal_destroy_rhdl_resource();
1032 static inline void flush_completed_wrs(struct t3_wq
*wq
, struct t3_cq
*cq
)
1034 struct t3_swsq
*sqp
;
1035 __u32 ptr
= wq
->sq_rptr
;
1036 int count
= Q_COUNT(wq
->sq_rptr
, wq
->sq_wptr
);
1038 sqp
= wq
->sq
+ Q_PTR2IDX(ptr
, wq
->sq_size_log2
);
1040 if (!sqp
->signaled
) {
1042 sqp
= wq
->sq
+ Q_PTR2IDX(ptr
, wq
->sq_size_log2
);
1043 } else if (sqp
->complete
) {
1046 * Insert this completed cqe into the swcq.
1048 PDBG("%s moving cqe into swcq sq idx %ld cq idx %ld\n",
1049 __FUNCTION__
, Q_PTR2IDX(ptr
, wq
->sq_size_log2
),
1050 Q_PTR2IDX(cq
->sw_wptr
, cq
->size_log2
));
1051 sqp
->cqe
.header
|= htonl(V_CQE_SWCQE(1));
1052 *(cq
->sw_queue
+ Q_PTR2IDX(cq
->sw_wptr
, cq
->size_log2
))
1061 static inline void create_read_req_cqe(struct t3_wq
*wq
,
1062 struct t3_cqe
*hw_cqe
,
1063 struct t3_cqe
*read_cqe
)
1065 read_cqe
->u
.scqe
.wrid_hi
= wq
->oldest_read
->sq_wptr
;
1066 read_cqe
->len
= wq
->oldest_read
->read_len
;
1067 read_cqe
->header
= htonl(V_CQE_QPID(CQE_QPID(*hw_cqe
)) |
1068 V_CQE_SWCQE(SW_CQE(*hw_cqe
)) |
1069 V_CQE_OPCODE(T3_READ_REQ
) |
1074 * Return a ptr to the next read wr in the SWSQ or NULL.
1076 static inline void advance_oldest_read(struct t3_wq
*wq
)
1079 u32 rptr
= wq
->oldest_read
- wq
->sq
+ 1;
1080 u32 wptr
= Q_PTR2IDX(wq
->sq_wptr
, wq
->sq_size_log2
);
1082 while (Q_PTR2IDX(rptr
, wq
->sq_size_log2
) != wptr
) {
1083 wq
->oldest_read
= wq
->sq
+ Q_PTR2IDX(rptr
, wq
->sq_size_log2
);
1085 if (wq
->oldest_read
->opcode
== T3_READ_REQ
)
1089 wq
->oldest_read
= NULL
;
1096 * check the validity of the first CQE,
1097 * supply the wq assicated with the qpid.
1099 * credit: cq credit to return to sge.
1100 * cqe_flushed: 1 iff the CQE is flushed.
1101 * cqe: copy of the polled CQE.
1105 * -1 CQE skipped, try again.
1107 int cxio_poll_cq(struct t3_wq
*wq
, struct t3_cq
*cq
, struct t3_cqe
*cqe
,
1108 u8
*cqe_flushed
, u64
*cookie
, u32
*credit
)
1111 struct t3_cqe
*hw_cqe
, read_cqe
;
1115 hw_cqe
= cxio_next_cqe(cq
);
1117 PDBG("%s CQE OOO %d qpid 0x%0x genbit %d type %d status 0x%0x"
1118 " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
1119 __FUNCTION__
, CQE_OOO(*hw_cqe
), CQE_QPID(*hw_cqe
),
1120 CQE_GENBIT(*hw_cqe
), CQE_TYPE(*hw_cqe
), CQE_STATUS(*hw_cqe
),
1121 CQE_OPCODE(*hw_cqe
), CQE_LEN(*hw_cqe
), CQE_WRID_HI(*hw_cqe
),
1122 CQE_WRID_LOW(*hw_cqe
));
1125 * skip cqe's not affiliated with a QP.
1133 * Gotta tweak READ completions:
1134 * 1) the cqe doesn't contain the sq_wptr from the wr.
1135 * 2) opcode not reflected from the wr.
1136 * 3) read_len not reflected from the wr.
1137 * 4) cq_type is RQ_TYPE not SQ_TYPE.
1139 if (RQ_TYPE(*hw_cqe
) && (CQE_OPCODE(*hw_cqe
) == T3_READ_RESP
)) {
1142 * Don't write to the HWCQ, so create a new read req CQE
1145 create_read_req_cqe(wq
, hw_cqe
, &read_cqe
);
1147 advance_oldest_read(wq
);
1151 * T3A: Discard TERMINATE CQEs.
1153 if (CQE_OPCODE(*hw_cqe
) == T3_TERMINATE
) {
1159 if (CQE_STATUS(*hw_cqe
) || wq
->error
) {
1160 *cqe_flushed
= wq
->error
;
1164 * T3A inserts errors into the CQE. We cannot return
1165 * these as work completions.
1167 /* incoming write failures */
1168 if ((CQE_OPCODE(*hw_cqe
) == T3_RDMA_WRITE
)
1169 && RQ_TYPE(*hw_cqe
)) {
1173 /* incoming read request failures */
1174 if ((CQE_OPCODE(*hw_cqe
) == T3_READ_RESP
) && SQ_TYPE(*hw_cqe
)) {
1179 /* incoming SEND with no receive posted failures */
1180 if ((CQE_OPCODE(*hw_cqe
) == T3_SEND
) && RQ_TYPE(*hw_cqe
) &&
1181 Q_EMPTY(wq
->rq_rptr
, wq
->rq_wptr
)) {
1191 if (RQ_TYPE(*hw_cqe
)) {
1194 * HW only validates 4 bits of MSN. So we must validate that
1195 * the MSN in the SEND is the next expected MSN. If its not,
1196 * then we complete this with TPT_ERR_MSN and mark the wq in
1199 if (unlikely((CQE_WRID_MSN(*hw_cqe
) != (wq
->rq_rptr
+ 1)))) {
1201 hw_cqe
->header
|= htonl(V_CQE_STATUS(TPT_ERR_MSN
));
1208 * If we get here its a send completion.
1210 * Handle out of order completion. These get stuffed
1211 * in the SW SQ. Then the SW SQ is walked to move any
1212 * now in-order completions into the SW CQ. This handles
1214 * 1) reaping unsignaled WRs when the first subsequent
1215 * signaled WR is completed.
1216 * 2) out of order read completions.
1218 if (!SW_CQE(*hw_cqe
) && (CQE_WRID_SQ_WPTR(*hw_cqe
) != wq
->sq_rptr
)) {
1219 struct t3_swsq
*sqp
;
1221 PDBG("%s out of order completion going in swsq at idx %ld\n",
1223 Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe
), wq
->sq_size_log2
));
1225 Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe
), wq
->sq_size_log2
);
1236 * Reap the associated WR(s) that are freed up with this
1239 if (SQ_TYPE(*hw_cqe
)) {
1240 wq
->sq_rptr
= CQE_WRID_SQ_WPTR(*hw_cqe
);
1241 PDBG("%s completing sq idx %ld\n", __FUNCTION__
,
1242 Q_PTR2IDX(wq
->sq_rptr
, wq
->sq_size_log2
));
1244 Q_PTR2IDX(wq
->sq_rptr
, wq
->sq_size_log2
))->wr_id
;
1247 PDBG("%s completing rq idx %ld\n", __FUNCTION__
,
1248 Q_PTR2IDX(wq
->rq_rptr
, wq
->rq_size_log2
));
1249 *cookie
= *(wq
->rq
+ Q_PTR2IDX(wq
->rq_rptr
, wq
->rq_size_log2
));
1255 * Flush any completed cqes that are now in-order.
1257 flush_completed_wrs(wq
, cq
);
1260 if (SW_CQE(*hw_cqe
)) {
1261 PDBG("%s cq %p cqid 0x%x skip sw cqe sw_rptr 0x%x\n",
1262 __FUNCTION__
, cq
, cq
->cqid
, cq
->sw_rptr
);
1265 PDBG("%s cq %p cqid 0x%x skip hw cqe rptr 0x%x\n",
1266 __FUNCTION__
, cq
, cq
->cqid
, cq
->rptr
);
1270 * T3A: compute credits.
1272 if (((cq
->rptr
- cq
->wptr
) > (1 << (cq
->size_log2
- 1)))
1273 || ((cq
->rptr
- cq
->wptr
) >= 128)) {
1274 *credit
= cq
->rptr
- cq
->wptr
;
1275 cq
->wptr
= cq
->rptr
;