RDMA/cxgb3: Wrap the software send queue pointer as needed on flush
[linux-2.6] / drivers / infiniband / hw / cxgb3 / cxio_hal.c
1 /*
2  * Copyright (c) 2006 Chelsio, Inc. All rights reserved.
3  *
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:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
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.
22  *
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
30  * SOFTWARE.
31  */
32 #include <asm/delay.h>
33
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>
39 #include <linux/dma-mapping.h>
40 #include <net/net_namespace.h>
41
42 #include "cxio_resource.h"
43 #include "cxio_hal.h"
44 #include "cxgb3_offload.h"
45 #include "sge_defs.h"
46
47 static LIST_HEAD(rdev_list);
48 static cxio_hal_ev_callback_func_t cxio_ev_cb = NULL;
49
50 static struct cxio_rdev *cxio_hal_find_rdev_by_name(char *dev_name)
51 {
52         struct cxio_rdev *rdev;
53
54         list_for_each_entry(rdev, &rdev_list, entry)
55                 if (!strcmp(rdev->dev_name, dev_name))
56                         return rdev;
57         return NULL;
58 }
59
60 static struct cxio_rdev *cxio_hal_find_rdev_by_t3cdev(struct t3cdev *tdev)
61 {
62         struct cxio_rdev *rdev;
63
64         list_for_each_entry(rdev, &rdev_list, entry)
65                 if (rdev->t3cdev_p == tdev)
66                         return rdev;
67         return NULL;
68 }
69
70 int cxio_hal_cq_op(struct cxio_rdev *rdev_p, struct t3_cq *cq,
71                    enum t3_cq_opcode op, u32 credit)
72 {
73         int ret;
74         struct t3_cqe *cqe;
75         u32 rptr;
76
77         struct rdma_cq_op setup;
78         setup.id = cq->cqid;
79         setup.credits = (op == CQ_CREDIT_UPDATE) ? credit : 0;
80         setup.op = op;
81         ret = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_OP, &setup);
82
83         if ((ret < 0) || (op == CQ_CREDIT_UPDATE))
84                 return ret;
85
86         /*
87          * If the rearm returned an index other than our current index,
88          * then there might be CQE's in flight (being DMA'd).  We must wait
89          * here for them to complete or the consumer can miss a notification.
90          */
91         if (Q_PTR2IDX((cq->rptr), cq->size_log2) != ret) {
92                 int i=0;
93
94                 rptr = cq->rptr;
95
96                 /*
97                  * Keep the generation correct by bumping rptr until it
98                  * matches the index returned by the rearm - 1.
99                  */
100                 while (Q_PTR2IDX((rptr+1), cq->size_log2) != ret)
101                         rptr++;
102
103                 /*
104                  * Now rptr is the index for the (last) cqe that was
105                  * in-flight at the time the HW rearmed the CQ.  We
106                  * spin until that CQE is valid.
107                  */
108                 cqe = cq->queue + Q_PTR2IDX(rptr, cq->size_log2);
109                 while (!CQ_VLD_ENTRY(rptr, cq->size_log2, cqe)) {
110                         udelay(1);
111                         if (i++ > 1000000) {
112                                 BUG_ON(1);
113                                 printk(KERN_ERR "%s: stalled rnic\n",
114                                        rdev_p->dev_name);
115                                 return -EIO;
116                         }
117                 }
118
119                 return 1;
120         }
121
122         return 0;
123 }
124
125 static int cxio_hal_clear_cq_ctx(struct cxio_rdev *rdev_p, u32 cqid)
126 {
127         struct rdma_cq_setup setup;
128         setup.id = cqid;
129         setup.base_addr = 0;    /* NULL address */
130         setup.size = 0;         /* disaable the CQ */
131         setup.credits = 0;
132         setup.credit_thres = 0;
133         setup.ovfl_mode = 0;
134         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
135 }
136
137 static int cxio_hal_clear_qp_ctx(struct cxio_rdev *rdev_p, u32 qpid)
138 {
139         u64 sge_cmd;
140         struct t3_modify_qp_wr *wqe;
141         struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
142         if (!skb) {
143                 PDBG("%s alloc_skb failed\n", __func__);
144                 return -ENOMEM;
145         }
146         wqe = (struct t3_modify_qp_wr *) skb_put(skb, sizeof(*wqe));
147         memset(wqe, 0, sizeof(*wqe));
148         build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD, 3, 0, qpid, 7);
149         wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
150         sge_cmd = qpid << 8 | 3;
151         wqe->sge_cmd = cpu_to_be64(sge_cmd);
152         skb->priority = CPL_PRIORITY_CONTROL;
153         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
154 }
155
156 int cxio_create_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
157 {
158         struct rdma_cq_setup setup;
159         int size = (1UL << (cq->size_log2)) * sizeof(struct t3_cqe);
160
161         cq->cqid = cxio_hal_get_cqid(rdev_p->rscp);
162         if (!cq->cqid)
163                 return -ENOMEM;
164         cq->sw_queue = kzalloc(size, GFP_KERNEL);
165         if (!cq->sw_queue)
166                 return -ENOMEM;
167         cq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev),
168                                              (1UL << (cq->size_log2)) *
169                                              sizeof(struct t3_cqe),
170                                              &(cq->dma_addr), GFP_KERNEL);
171         if (!cq->queue) {
172                 kfree(cq->sw_queue);
173                 return -ENOMEM;
174         }
175         pci_unmap_addr_set(cq, mapping, cq->dma_addr);
176         memset(cq->queue, 0, size);
177         setup.id = cq->cqid;
178         setup.base_addr = (u64) (cq->dma_addr);
179         setup.size = 1UL << cq->size_log2;
180         setup.credits = 65535;
181         setup.credit_thres = 1;
182         if (rdev_p->t3cdev_p->type != T3A)
183                 setup.ovfl_mode = 0;
184         else
185                 setup.ovfl_mode = 1;
186         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
187 }
188
189 int cxio_resize_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
190 {
191         struct rdma_cq_setup setup;
192         setup.id = cq->cqid;
193         setup.base_addr = (u64) (cq->dma_addr);
194         setup.size = 1UL << cq->size_log2;
195         setup.credits = setup.size;
196         setup.credit_thres = setup.size;        /* TBD: overflow recovery */
197         setup.ovfl_mode = 1;
198         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
199 }
200
201 static u32 get_qpid(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
202 {
203         struct cxio_qpid_list *entry;
204         u32 qpid;
205         int i;
206
207         mutex_lock(&uctx->lock);
208         if (!list_empty(&uctx->qpids)) {
209                 entry = list_entry(uctx->qpids.next, struct cxio_qpid_list,
210                                    entry);
211                 list_del(&entry->entry);
212                 qpid = entry->qpid;
213                 kfree(entry);
214         } else {
215                 qpid = cxio_hal_get_qpid(rdev_p->rscp);
216                 if (!qpid)
217                         goto out;
218                 for (i = qpid+1; i & rdev_p->qpmask; i++) {
219                         entry = kmalloc(sizeof *entry, GFP_KERNEL);
220                         if (!entry)
221                                 break;
222                         entry->qpid = i;
223                         list_add_tail(&entry->entry, &uctx->qpids);
224                 }
225         }
226 out:
227         mutex_unlock(&uctx->lock);
228         PDBG("%s qpid 0x%x\n", __func__, qpid);
229         return qpid;
230 }
231
232 static void put_qpid(struct cxio_rdev *rdev_p, u32 qpid,
233                      struct cxio_ucontext *uctx)
234 {
235         struct cxio_qpid_list *entry;
236
237         entry = kmalloc(sizeof *entry, GFP_KERNEL);
238         if (!entry)
239                 return;
240         PDBG("%s qpid 0x%x\n", __func__, qpid);
241         entry->qpid = qpid;
242         mutex_lock(&uctx->lock);
243         list_add_tail(&entry->entry, &uctx->qpids);
244         mutex_unlock(&uctx->lock);
245 }
246
247 void cxio_release_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
248 {
249         struct list_head *pos, *nxt;
250         struct cxio_qpid_list *entry;
251
252         mutex_lock(&uctx->lock);
253         list_for_each_safe(pos, nxt, &uctx->qpids) {
254                 entry = list_entry(pos, struct cxio_qpid_list, entry);
255                 list_del_init(&entry->entry);
256                 if (!(entry->qpid & rdev_p->qpmask))
257                         cxio_hal_put_qpid(rdev_p->rscp, entry->qpid);
258                 kfree(entry);
259         }
260         mutex_unlock(&uctx->lock);
261 }
262
263 void cxio_init_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
264 {
265         INIT_LIST_HEAD(&uctx->qpids);
266         mutex_init(&uctx->lock);
267 }
268
269 int cxio_create_qp(struct cxio_rdev *rdev_p, u32 kernel_domain,
270                    struct t3_wq *wq, struct cxio_ucontext *uctx)
271 {
272         int depth = 1UL << wq->size_log2;
273         int rqsize = 1UL << wq->rq_size_log2;
274
275         wq->qpid = get_qpid(rdev_p, uctx);
276         if (!wq->qpid)
277                 return -ENOMEM;
278
279         wq->rq = kzalloc(depth * sizeof(u64), GFP_KERNEL);
280         if (!wq->rq)
281                 goto err1;
282
283         wq->rq_addr = cxio_hal_rqtpool_alloc(rdev_p, rqsize);
284         if (!wq->rq_addr)
285                 goto err2;
286
287         wq->sq = kzalloc(depth * sizeof(struct t3_swsq), GFP_KERNEL);
288         if (!wq->sq)
289                 goto err3;
290
291         wq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev),
292                                              depth * sizeof(union t3_wr),
293                                              &(wq->dma_addr), GFP_KERNEL);
294         if (!wq->queue)
295                 goto err4;
296
297         memset(wq->queue, 0, depth * sizeof(union t3_wr));
298         pci_unmap_addr_set(wq, mapping, wq->dma_addr);
299         wq->doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
300         if (!kernel_domain)
301                 wq->udb = (u64)rdev_p->rnic_info.udbell_physbase +
302                                         (wq->qpid << rdev_p->qpshift);
303         PDBG("%s qpid 0x%x doorbell 0x%p udb 0x%llx\n", __func__,
304              wq->qpid, wq->doorbell, (unsigned long long) wq->udb);
305         return 0;
306 err4:
307         kfree(wq->sq);
308 err3:
309         cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, rqsize);
310 err2:
311         kfree(wq->rq);
312 err1:
313         put_qpid(rdev_p, wq->qpid, uctx);
314         return -ENOMEM;
315 }
316
317 int cxio_destroy_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
318 {
319         int err;
320         err = cxio_hal_clear_cq_ctx(rdev_p, cq->cqid);
321         kfree(cq->sw_queue);
322         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
323                           (1UL << (cq->size_log2))
324                           * sizeof(struct t3_cqe), cq->queue,
325                           pci_unmap_addr(cq, mapping));
326         cxio_hal_put_cqid(rdev_p->rscp, cq->cqid);
327         return err;
328 }
329
330 int cxio_destroy_qp(struct cxio_rdev *rdev_p, struct t3_wq *wq,
331                     struct cxio_ucontext *uctx)
332 {
333         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
334                           (1UL << (wq->size_log2))
335                           * sizeof(union t3_wr), wq->queue,
336                           pci_unmap_addr(wq, mapping));
337         kfree(wq->sq);
338         cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, (1UL << wq->rq_size_log2));
339         kfree(wq->rq);
340         put_qpid(rdev_p, wq->qpid, uctx);
341         return 0;
342 }
343
344 static void insert_recv_cqe(struct t3_wq *wq, struct t3_cq *cq)
345 {
346         struct t3_cqe cqe;
347
348         PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __func__,
349              wq, cq, cq->sw_rptr, cq->sw_wptr);
350         memset(&cqe, 0, sizeof(cqe));
351         cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
352                                  V_CQE_OPCODE(T3_SEND) |
353                                  V_CQE_TYPE(0) |
354                                  V_CQE_SWCQE(1) |
355                                  V_CQE_QPID(wq->qpid) |
356                                  V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
357                                                        cq->size_log2)));
358         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
359         cq->sw_wptr++;
360 }
361
362 int cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count)
363 {
364         u32 ptr;
365         int flushed = 0;
366
367         PDBG("%s wq %p cq %p\n", __func__, wq, cq);
368
369         /* flush RQ */
370         PDBG("%s rq_rptr %u rq_wptr %u skip count %u\n", __func__,
371             wq->rq_rptr, wq->rq_wptr, count);
372         ptr = wq->rq_rptr + count;
373         while (ptr++ != wq->rq_wptr) {
374                 insert_recv_cqe(wq, cq);
375                 flushed++;
376         }
377         return flushed;
378 }
379
380 static void insert_sq_cqe(struct t3_wq *wq, struct t3_cq *cq,
381                           struct t3_swsq *sqp)
382 {
383         struct t3_cqe cqe;
384
385         PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __func__,
386              wq, cq, cq->sw_rptr, cq->sw_wptr);
387         memset(&cqe, 0, sizeof(cqe));
388         cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
389                                  V_CQE_OPCODE(sqp->opcode) |
390                                  V_CQE_TYPE(1) |
391                                  V_CQE_SWCQE(1) |
392                                  V_CQE_QPID(wq->qpid) |
393                                  V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
394                                                        cq->size_log2)));
395         cqe.u.scqe.wrid_hi = sqp->sq_wptr;
396
397         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
398         cq->sw_wptr++;
399 }
400
401 int cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count)
402 {
403         __u32 ptr;
404         int flushed = 0;
405         struct t3_swsq *sqp = wq->sq + Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2);
406
407         ptr = wq->sq_rptr + count;
408         sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
409         while (ptr != wq->sq_wptr) {
410                 insert_sq_cqe(wq, cq, sqp);
411                 ptr++;
412                 sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
413                 flushed++;
414         }
415         return flushed;
416 }
417
418 /*
419  * Move all CQEs from the HWCQ into the SWCQ.
420  */
421 void cxio_flush_hw_cq(struct t3_cq *cq)
422 {
423         struct t3_cqe *cqe, *swcqe;
424
425         PDBG("%s cq %p cqid 0x%x\n", __func__, cq, cq->cqid);
426         cqe = cxio_next_hw_cqe(cq);
427         while (cqe) {
428                 PDBG("%s flushing hwcq rptr 0x%x to swcq wptr 0x%x\n",
429                      __func__, cq->rptr, cq->sw_wptr);
430                 swcqe = cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2);
431                 *swcqe = *cqe;
432                 swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
433                 cq->sw_wptr++;
434                 cq->rptr++;
435                 cqe = cxio_next_hw_cqe(cq);
436         }
437 }
438
439 static int cqe_completes_wr(struct t3_cqe *cqe, struct t3_wq *wq)
440 {
441         if (CQE_OPCODE(*cqe) == T3_TERMINATE)
442                 return 0;
443
444         if ((CQE_OPCODE(*cqe) == T3_RDMA_WRITE) && RQ_TYPE(*cqe))
445                 return 0;
446
447         if ((CQE_OPCODE(*cqe) == T3_READ_RESP) && SQ_TYPE(*cqe))
448                 return 0;
449
450         if ((CQE_OPCODE(*cqe) == T3_SEND) && RQ_TYPE(*cqe) &&
451             Q_EMPTY(wq->rq_rptr, wq->rq_wptr))
452                 return 0;
453
454         return 1;
455 }
456
457 void cxio_count_scqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
458 {
459         struct t3_cqe *cqe;
460         u32 ptr;
461
462         *count = 0;
463         ptr = cq->sw_rptr;
464         while (!Q_EMPTY(ptr, cq->sw_wptr)) {
465                 cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
466                 if ((SQ_TYPE(*cqe) ||
467                      ((CQE_OPCODE(*cqe) == T3_READ_RESP) && wq->oldest_read)) &&
468                     (CQE_QPID(*cqe) == wq->qpid))
469                         (*count)++;
470                 ptr++;
471         }
472         PDBG("%s cq %p count %d\n", __func__, cq, *count);
473 }
474
475 void cxio_count_rcqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
476 {
477         struct t3_cqe *cqe;
478         u32 ptr;
479
480         *count = 0;
481         PDBG("%s count zero %d\n", __func__, *count);
482         ptr = cq->sw_rptr;
483         while (!Q_EMPTY(ptr, cq->sw_wptr)) {
484                 cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
485                 if (RQ_TYPE(*cqe) && (CQE_OPCODE(*cqe) != T3_READ_RESP) &&
486                     (CQE_QPID(*cqe) == wq->qpid) && cqe_completes_wr(cqe, wq))
487                         (*count)++;
488                 ptr++;
489         }
490         PDBG("%s cq %p count %d\n", __func__, cq, *count);
491 }
492
493 static int cxio_hal_init_ctrl_cq(struct cxio_rdev *rdev_p)
494 {
495         struct rdma_cq_setup setup;
496         setup.id = 0;
497         setup.base_addr = 0;    /* NULL address */
498         setup.size = 1;         /* enable the CQ */
499         setup.credits = 0;
500
501         /* force SGE to redirect to RspQ and interrupt */
502         setup.credit_thres = 0;
503         setup.ovfl_mode = 1;
504         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
505 }
506
507 static int cxio_hal_init_ctrl_qp(struct cxio_rdev *rdev_p)
508 {
509         int err;
510         u64 sge_cmd, ctx0, ctx1;
511         u64 base_addr;
512         struct t3_modify_qp_wr *wqe;
513         struct sk_buff *skb;
514
515         skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
516         if (!skb) {
517                 PDBG("%s alloc_skb failed\n", __func__);
518                 return -ENOMEM;
519         }
520         err = cxio_hal_init_ctrl_cq(rdev_p);
521         if (err) {
522                 PDBG("%s err %d initializing ctrl_cq\n", __func__, err);
523                 goto err;
524         }
525         rdev_p->ctrl_qp.workq = dma_alloc_coherent(
526                                         &(rdev_p->rnic_info.pdev->dev),
527                                         (1 << T3_CTRL_QP_SIZE_LOG2) *
528                                         sizeof(union t3_wr),
529                                         &(rdev_p->ctrl_qp.dma_addr),
530                                         GFP_KERNEL);
531         if (!rdev_p->ctrl_qp.workq) {
532                 PDBG("%s dma_alloc_coherent failed\n", __func__);
533                 err = -ENOMEM;
534                 goto err;
535         }
536         pci_unmap_addr_set(&rdev_p->ctrl_qp, mapping,
537                            rdev_p->ctrl_qp.dma_addr);
538         rdev_p->ctrl_qp.doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
539         memset(rdev_p->ctrl_qp.workq, 0,
540                (1 << T3_CTRL_QP_SIZE_LOG2) * sizeof(union t3_wr));
541
542         mutex_init(&rdev_p->ctrl_qp.lock);
543         init_waitqueue_head(&rdev_p->ctrl_qp.waitq);
544
545         /* update HW Ctrl QP context */
546         base_addr = rdev_p->ctrl_qp.dma_addr;
547         base_addr >>= 12;
548         ctx0 = (V_EC_SIZE((1 << T3_CTRL_QP_SIZE_LOG2)) |
549                 V_EC_BASE_LO((u32) base_addr & 0xffff));
550         ctx0 <<= 32;
551         ctx0 |= V_EC_CREDITS(FW_WR_NUM);
552         base_addr >>= 16;
553         ctx1 = (u32) base_addr;
554         base_addr >>= 32;
555         ctx1 |= ((u64) (V_EC_BASE_HI((u32) base_addr & 0xf) | V_EC_RESPQ(0) |
556                         V_EC_TYPE(0) | V_EC_GEN(1) |
557                         V_EC_UP_TOKEN(T3_CTL_QP_TID) | F_EC_VALID)) << 32;
558         wqe = (struct t3_modify_qp_wr *) skb_put(skb, sizeof(*wqe));
559         memset(wqe, 0, sizeof(*wqe));
560         build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD, 0, 0,
561                        T3_CTL_QP_TID, 7);
562         wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
563         sge_cmd = (3ULL << 56) | FW_RI_SGEEC_START << 8 | 3;
564         wqe->sge_cmd = cpu_to_be64(sge_cmd);
565         wqe->ctx1 = cpu_to_be64(ctx1);
566         wqe->ctx0 = cpu_to_be64(ctx0);
567         PDBG("CtrlQP dma_addr 0x%llx workq %p size %d\n",
568              (unsigned long long) rdev_p->ctrl_qp.dma_addr,
569              rdev_p->ctrl_qp.workq, 1 << T3_CTRL_QP_SIZE_LOG2);
570         skb->priority = CPL_PRIORITY_CONTROL;
571         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
572 err:
573         kfree_skb(skb);
574         return err;
575 }
576
577 static int cxio_hal_destroy_ctrl_qp(struct cxio_rdev *rdev_p)
578 {
579         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
580                           (1UL << T3_CTRL_QP_SIZE_LOG2)
581                           * sizeof(union t3_wr), rdev_p->ctrl_qp.workq,
582                           pci_unmap_addr(&rdev_p->ctrl_qp, mapping));
583         return cxio_hal_clear_qp_ctx(rdev_p, T3_CTRL_QP_ID);
584 }
585
586 /* write len bytes of data into addr (32B aligned address)
587  * If data is NULL, clear len byte of memory to zero.
588  * caller aquires the ctrl_qp lock before the call
589  */
590 static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev *rdev_p, u32 addr,
591                                       u32 len, void *data)
592 {
593         u32 i, nr_wqe, copy_len;
594         u8 *copy_data;
595         u8 wr_len, utx_len;     /* length in 8 byte flit */
596         enum t3_wr_flags flag;
597         __be64 *wqe;
598         u64 utx_cmd;
599         addr &= 0x7FFFFFF;
600         nr_wqe = len % 96 ? len / 96 + 1 : len / 96;    /* 96B max per WQE */
601         PDBG("%s wptr 0x%x rptr 0x%x len %d, nr_wqe %d data %p addr 0x%0x\n",
602              __func__, rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, len,
603              nr_wqe, data, addr);
604         utx_len = 3;            /* in 32B unit */
605         for (i = 0; i < nr_wqe; i++) {
606                 if (Q_FULL(rdev_p->ctrl_qp.rptr, rdev_p->ctrl_qp.wptr,
607                            T3_CTRL_QP_SIZE_LOG2)) {
608                         PDBG("%s ctrl_qp full wtpr 0x%0x rptr 0x%0x, "
609                              "wait for more space i %d\n", __func__,
610                              rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, i);
611                         if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
612                                              !Q_FULL(rdev_p->ctrl_qp.rptr,
613                                                      rdev_p->ctrl_qp.wptr,
614                                                      T3_CTRL_QP_SIZE_LOG2))) {
615                                 PDBG("%s ctrl_qp workq interrupted\n",
616                                      __func__);
617                                 return -ERESTARTSYS;
618                         }
619                         PDBG("%s ctrl_qp wakeup, continue posting work request "
620                              "i %d\n", __func__, i);
621                 }
622                 wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
623                                                 (1 << T3_CTRL_QP_SIZE_LOG2)));
624                 flag = 0;
625                 if (i == (nr_wqe - 1)) {
626                         /* last WQE */
627                         flag = T3_COMPLETION_FLAG;
628                         if (len % 32)
629                                 utx_len = len / 32 + 1;
630                         else
631                                 utx_len = len / 32;
632                 }
633
634                 /*
635                  * Force a CQE to return the credit to the workq in case
636                  * we posted more than half the max QP size of WRs
637                  */
638                 if ((i != 0) &&
639                     (i % (((1 << T3_CTRL_QP_SIZE_LOG2)) >> 1) == 0)) {
640                         flag = T3_COMPLETION_FLAG;
641                         PDBG("%s force completion at i %d\n", __func__, i);
642                 }
643
644                 /* build the utx mem command */
645                 wqe += (sizeof(struct t3_bypass_wr) >> 3);
646                 utx_cmd = (T3_UTX_MEM_WRITE << 28) | (addr + i * 3);
647                 utx_cmd <<= 32;
648                 utx_cmd |= (utx_len << 28) | ((utx_len << 2) + 1);
649                 *wqe = cpu_to_be64(utx_cmd);
650                 wqe++;
651                 copy_data = (u8 *) data + i * 96;
652                 copy_len = len > 96 ? 96 : len;
653
654                 /* clear memory content if data is NULL */
655                 if (data)
656                         memcpy(wqe, copy_data, copy_len);
657                 else
658                         memset(wqe, 0, copy_len);
659                 if (copy_len % 32)
660                         memset(((u8 *) wqe) + copy_len, 0,
661                                32 - (copy_len % 32));
662                 wr_len = ((sizeof(struct t3_bypass_wr)) >> 3) + 1 +
663                          (utx_len << 2);
664                 wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
665                               (1 << T3_CTRL_QP_SIZE_LOG2)));
666
667                 /* wptr in the WRID[31:0] */
668                 ((union t3_wrid *)(wqe+1))->id0.low = rdev_p->ctrl_qp.wptr;
669
670                 /*
671                  * This must be the last write with a memory barrier
672                  * for the genbit
673                  */
674                 build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_BP, flag,
675                                Q_GENBIT(rdev_p->ctrl_qp.wptr,
676                                         T3_CTRL_QP_SIZE_LOG2), T3_CTRL_QP_ID,
677                                wr_len);
678                 if (flag == T3_COMPLETION_FLAG)
679                         ring_doorbell(rdev_p->ctrl_qp.doorbell, T3_CTRL_QP_ID);
680                 len -= 96;
681                 rdev_p->ctrl_qp.wptr++;
682         }
683         return 0;
684 }
685
686 /* IN: stag key, pdid, perm, zbva, to, len, page_size, pbl_size and pbl_addr
687  * OUT: stag index
688  * TBD: shared memory region support
689  */
690 static int __cxio_tpt_op(struct cxio_rdev *rdev_p, u32 reset_tpt_entry,
691                          u32 *stag, u8 stag_state, u32 pdid,
692                          enum tpt_mem_type type, enum tpt_mem_perm perm,
693                          u32 zbva, u64 to, u32 len, u8 page_size,
694                          u32 pbl_size, u32 pbl_addr)
695 {
696         int err;
697         struct tpt_entry tpt;
698         u32 stag_idx;
699         u32 wptr;
700
701         stag_state = stag_state > 0;
702         stag_idx = (*stag) >> 8;
703
704         if ((!reset_tpt_entry) && !(*stag != T3_STAG_UNSET)) {
705                 stag_idx = cxio_hal_get_stag(rdev_p->rscp);
706                 if (!stag_idx)
707                         return -ENOMEM;
708                 *stag = (stag_idx << 8) | ((*stag) & 0xFF);
709         }
710         PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
711              __func__, stag_state, type, pdid, stag_idx);
712
713         mutex_lock(&rdev_p->ctrl_qp.lock);
714
715         /* write TPT entry */
716         if (reset_tpt_entry)
717                 memset(&tpt, 0, sizeof(tpt));
718         else {
719                 tpt.valid_stag_pdid = cpu_to_be32(F_TPT_VALID |
720                                 V_TPT_STAG_KEY((*stag) & M_TPT_STAG_KEY) |
721                                 V_TPT_STAG_STATE(stag_state) |
722                                 V_TPT_STAG_TYPE(type) | V_TPT_PDID(pdid));
723                 BUG_ON(page_size >= 28);
724                 tpt.flags_pagesize_qpid = cpu_to_be32(V_TPT_PERM(perm) |
725                                 F_TPT_MW_BIND_ENABLE |
726                                 V_TPT_ADDR_TYPE((zbva ? TPT_ZBTO : TPT_VATO)) |
727                                 V_TPT_PAGE_SIZE(page_size));
728                 tpt.rsvd_pbl_addr = reset_tpt_entry ? 0 :
729                                     cpu_to_be32(V_TPT_PBL_ADDR(PBL_OFF(rdev_p, pbl_addr)>>3));
730                 tpt.len = cpu_to_be32(len);
731                 tpt.va_hi = cpu_to_be32((u32) (to >> 32));
732                 tpt.va_low_or_fbo = cpu_to_be32((u32) (to & 0xFFFFFFFFULL));
733                 tpt.rsvd_bind_cnt_or_pstag = 0;
734                 tpt.rsvd_pbl_size = reset_tpt_entry ? 0 :
735                                   cpu_to_be32(V_TPT_PBL_SIZE(pbl_size >> 2));
736         }
737         err = cxio_hal_ctrl_qp_write_mem(rdev_p,
738                                        stag_idx +
739                                        (rdev_p->rnic_info.tpt_base >> 5),
740                                        sizeof(tpt), &tpt);
741
742         /* release the stag index to free pool */
743         if (reset_tpt_entry)
744                 cxio_hal_put_stag(rdev_p->rscp, stag_idx);
745
746         wptr = rdev_p->ctrl_qp.wptr;
747         mutex_unlock(&rdev_p->ctrl_qp.lock);
748         if (!err)
749                 if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
750                                              SEQ32_GE(rdev_p->ctrl_qp.rptr,
751                                                       wptr)))
752                         return -ERESTARTSYS;
753         return err;
754 }
755
756 int cxio_write_pbl(struct cxio_rdev *rdev_p, __be64 *pbl,
757                    u32 pbl_addr, u32 pbl_size)
758 {
759         u32 wptr;
760         int err;
761
762         PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
763              __func__, pbl_addr, rdev_p->rnic_info.pbl_base,
764              pbl_size);
765
766         mutex_lock(&rdev_p->ctrl_qp.lock);
767         err = cxio_hal_ctrl_qp_write_mem(rdev_p, pbl_addr >> 5, pbl_size << 3,
768                                          pbl);
769         wptr = rdev_p->ctrl_qp.wptr;
770         mutex_unlock(&rdev_p->ctrl_qp.lock);
771         if (err)
772                 return err;
773
774         if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
775                                      SEQ32_GE(rdev_p->ctrl_qp.rptr,
776                                               wptr)))
777                 return -ERESTARTSYS;
778
779         return 0;
780 }
781
782 int cxio_register_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
783                            enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
784                            u8 page_size, u32 pbl_size, u32 pbl_addr)
785 {
786         *stag = T3_STAG_UNSET;
787         return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
788                              zbva, to, len, page_size, pbl_size, pbl_addr);
789 }
790
791 int cxio_reregister_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
792                            enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
793                            u8 page_size, u32 pbl_size, u32 pbl_addr)
794 {
795         return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
796                              zbva, to, len, page_size, pbl_size, pbl_addr);
797 }
798
799 int cxio_dereg_mem(struct cxio_rdev *rdev_p, u32 stag, u32 pbl_size,
800                    u32 pbl_addr)
801 {
802         return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
803                              pbl_size, pbl_addr);
804 }
805
806 int cxio_allocate_window(struct cxio_rdev *rdev_p, u32 * stag, u32 pdid)
807 {
808         *stag = T3_STAG_UNSET;
809         return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_MW, 0, 0, 0ULL, 0, 0,
810                              0, 0);
811 }
812
813 int cxio_deallocate_window(struct cxio_rdev *rdev_p, u32 stag)
814 {
815         return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
816                              0, 0);
817 }
818
819 int cxio_rdma_init(struct cxio_rdev *rdev_p, struct t3_rdma_init_attr *attr)
820 {
821         struct t3_rdma_init_wr *wqe;
822         struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_ATOMIC);
823         if (!skb)
824                 return -ENOMEM;
825         PDBG("%s rdev_p %p\n", __func__, rdev_p);
826         wqe = (struct t3_rdma_init_wr *) __skb_put(skb, sizeof(*wqe));
827         wqe->wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_INIT));
828         wqe->wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(attr->tid) |
829                                            V_FW_RIWR_LEN(sizeof(*wqe) >> 3));
830         wqe->wrid.id1 = 0;
831         wqe->qpid = cpu_to_be32(attr->qpid);
832         wqe->pdid = cpu_to_be32(attr->pdid);
833         wqe->scqid = cpu_to_be32(attr->scqid);
834         wqe->rcqid = cpu_to_be32(attr->rcqid);
835         wqe->rq_addr = cpu_to_be32(attr->rq_addr - rdev_p->rnic_info.rqt_base);
836         wqe->rq_size = cpu_to_be32(attr->rq_size);
837         wqe->mpaattrs = attr->mpaattrs;
838         wqe->qpcaps = attr->qpcaps;
839         wqe->ulpdu_size = cpu_to_be16(attr->tcp_emss);
840         wqe->rqe_count = cpu_to_be16(attr->rqe_count);
841         wqe->flags_rtr_type = cpu_to_be16(attr->flags|V_RTR_TYPE(attr->rtr_type));
842         wqe->ord = cpu_to_be32(attr->ord);
843         wqe->ird = cpu_to_be32(attr->ird);
844         wqe->qp_dma_addr = cpu_to_be64(attr->qp_dma_addr);
845         wqe->qp_dma_size = cpu_to_be32(attr->qp_dma_size);
846         wqe->irs = cpu_to_be32(attr->irs);
847         skb->priority = 0;      /* 0=>ToeQ; 1=>CtrlQ */
848         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
849 }
850
851 void cxio_register_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
852 {
853         cxio_ev_cb = ev_cb;
854 }
855
856 void cxio_unregister_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
857 {
858         cxio_ev_cb = NULL;
859 }
860
861 static int cxio_hal_ev_handler(struct t3cdev *t3cdev_p, struct sk_buff *skb)
862 {
863         static int cnt;
864         struct cxio_rdev *rdev_p = NULL;
865         struct respQ_msg_t *rsp_msg = (struct respQ_msg_t *) skb->data;
866         PDBG("%d: %s cq_id 0x%x cq_ptr 0x%x genbit %0x overflow %0x an %0x"
867              " se %0x notify %0x cqbranch %0x creditth %0x\n",
868              cnt, __func__, RSPQ_CQID(rsp_msg), RSPQ_CQPTR(rsp_msg),
869              RSPQ_GENBIT(rsp_msg), RSPQ_OVERFLOW(rsp_msg), RSPQ_AN(rsp_msg),
870              RSPQ_SE(rsp_msg), RSPQ_NOTIFY(rsp_msg), RSPQ_CQBRANCH(rsp_msg),
871              RSPQ_CREDIT_THRESH(rsp_msg));
872         PDBG("CQE: QPID 0x%0x genbit %0x type 0x%0x status 0x%0x opcode %d "
873              "len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
874              CQE_QPID(rsp_msg->cqe), CQE_GENBIT(rsp_msg->cqe),
875              CQE_TYPE(rsp_msg->cqe), CQE_STATUS(rsp_msg->cqe),
876              CQE_OPCODE(rsp_msg->cqe), CQE_LEN(rsp_msg->cqe),
877              CQE_WRID_HI(rsp_msg->cqe), CQE_WRID_LOW(rsp_msg->cqe));
878         rdev_p = (struct cxio_rdev *)t3cdev_p->ulp;
879         if (!rdev_p) {
880                 PDBG("%s called by t3cdev %p with null ulp\n", __func__,
881                      t3cdev_p);
882                 return 0;
883         }
884         if (CQE_QPID(rsp_msg->cqe) == T3_CTRL_QP_ID) {
885                 rdev_p->ctrl_qp.rptr = CQE_WRID_LOW(rsp_msg->cqe) + 1;
886                 wake_up_interruptible(&rdev_p->ctrl_qp.waitq);
887                 dev_kfree_skb_irq(skb);
888         } else if (CQE_QPID(rsp_msg->cqe) == 0xfff8)
889                 dev_kfree_skb_irq(skb);
890         else if (cxio_ev_cb)
891                 (*cxio_ev_cb) (rdev_p, skb);
892         else
893                 dev_kfree_skb_irq(skb);
894         cnt++;
895         return 0;
896 }
897
898 /* Caller takes care of locking if needed */
899 int cxio_rdev_open(struct cxio_rdev *rdev_p)
900 {
901         struct net_device *netdev_p = NULL;
902         int err = 0;
903         if (strlen(rdev_p->dev_name)) {
904                 if (cxio_hal_find_rdev_by_name(rdev_p->dev_name)) {
905                         return -EBUSY;
906                 }
907                 netdev_p = dev_get_by_name(&init_net, rdev_p->dev_name);
908                 if (!netdev_p) {
909                         return -EINVAL;
910                 }
911                 dev_put(netdev_p);
912         } else if (rdev_p->t3cdev_p) {
913                 if (cxio_hal_find_rdev_by_t3cdev(rdev_p->t3cdev_p)) {
914                         return -EBUSY;
915                 }
916                 netdev_p = rdev_p->t3cdev_p->lldev;
917                 strncpy(rdev_p->dev_name, rdev_p->t3cdev_p->name,
918                         T3_MAX_DEV_NAME_LEN);
919         } else {
920                 PDBG("%s t3cdev_p or dev_name must be set\n", __func__);
921                 return -EINVAL;
922         }
923
924         list_add_tail(&rdev_p->entry, &rdev_list);
925
926         PDBG("%s opening rnic dev %s\n", __func__, rdev_p->dev_name);
927         memset(&rdev_p->ctrl_qp, 0, sizeof(rdev_p->ctrl_qp));
928         if (!rdev_p->t3cdev_p)
929                 rdev_p->t3cdev_p = dev2t3cdev(netdev_p);
930         rdev_p->t3cdev_p->ulp = (void *) rdev_p;
931         err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_GET_PARAMS,
932                                          &(rdev_p->rnic_info));
933         if (err) {
934                 printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
935                      __func__, rdev_p->t3cdev_p, err);
936                 goto err1;
937         }
938         err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, GET_PORTS,
939                                     &(rdev_p->port_info));
940         if (err) {
941                 printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
942                      __func__, rdev_p->t3cdev_p, err);
943                 goto err1;
944         }
945
946         /*
947          * qpshift is the number of bits to shift the qpid left in order
948          * to get the correct address of the doorbell for that qp.
949          */
950         cxio_init_ucontext(rdev_p, &rdev_p->uctx);
951         rdev_p->qpshift = PAGE_SHIFT -
952                           ilog2(65536 >>
953                                     ilog2(rdev_p->rnic_info.udbell_len >>
954                                               PAGE_SHIFT));
955         rdev_p->qpnr = rdev_p->rnic_info.udbell_len >> PAGE_SHIFT;
956         rdev_p->qpmask = (65536 >> ilog2(rdev_p->qpnr)) - 1;
957         PDBG("%s rnic %s info: tpt_base 0x%0x tpt_top 0x%0x num stags %d "
958              "pbl_base 0x%0x pbl_top 0x%0x rqt_base 0x%0x, rqt_top 0x%0x\n",
959              __func__, rdev_p->dev_name, rdev_p->rnic_info.tpt_base,
960              rdev_p->rnic_info.tpt_top, cxio_num_stags(rdev_p),
961              rdev_p->rnic_info.pbl_base,
962              rdev_p->rnic_info.pbl_top, rdev_p->rnic_info.rqt_base,
963              rdev_p->rnic_info.rqt_top);
964         PDBG("udbell_len 0x%0x udbell_physbase 0x%lx kdb_addr %p qpshift %lu "
965              "qpnr %d qpmask 0x%x\n",
966              rdev_p->rnic_info.udbell_len,
967              rdev_p->rnic_info.udbell_physbase, rdev_p->rnic_info.kdb_addr,
968              rdev_p->qpshift, rdev_p->qpnr, rdev_p->qpmask);
969
970         err = cxio_hal_init_ctrl_qp(rdev_p);
971         if (err) {
972                 printk(KERN_ERR "%s error %d initializing ctrl_qp.\n",
973                        __func__, err);
974                 goto err1;
975         }
976         err = cxio_hal_init_resource(rdev_p, cxio_num_stags(rdev_p), 0,
977                                      0, T3_MAX_NUM_QP, T3_MAX_NUM_CQ,
978                                      T3_MAX_NUM_PD);
979         if (err) {
980                 printk(KERN_ERR "%s error %d initializing hal resources.\n",
981                        __func__, err);
982                 goto err2;
983         }
984         err = cxio_hal_pblpool_create(rdev_p);
985         if (err) {
986                 printk(KERN_ERR "%s error %d initializing pbl mem pool.\n",
987                        __func__, err);
988                 goto err3;
989         }
990         err = cxio_hal_rqtpool_create(rdev_p);
991         if (err) {
992                 printk(KERN_ERR "%s error %d initializing rqt mem pool.\n",
993                        __func__, err);
994                 goto err4;
995         }
996         return 0;
997 err4:
998         cxio_hal_pblpool_destroy(rdev_p);
999 err3:
1000         cxio_hal_destroy_resource(rdev_p->rscp);
1001 err2:
1002         cxio_hal_destroy_ctrl_qp(rdev_p);
1003 err1:
1004         list_del(&rdev_p->entry);
1005         return err;
1006 }
1007
1008 void cxio_rdev_close(struct cxio_rdev *rdev_p)
1009 {
1010         if (rdev_p) {
1011                 cxio_hal_pblpool_destroy(rdev_p);
1012                 cxio_hal_rqtpool_destroy(rdev_p);
1013                 list_del(&rdev_p->entry);
1014                 rdev_p->t3cdev_p->ulp = NULL;
1015                 cxio_hal_destroy_ctrl_qp(rdev_p);
1016                 cxio_hal_destroy_resource(rdev_p->rscp);
1017         }
1018 }
1019
1020 int __init cxio_hal_init(void)
1021 {
1022         if (cxio_hal_init_rhdl_resource(T3_MAX_NUM_RI))
1023                 return -ENOMEM;
1024         t3_register_cpl_handler(CPL_ASYNC_NOTIF, cxio_hal_ev_handler);
1025         return 0;
1026 }
1027
1028 void __exit cxio_hal_exit(void)
1029 {
1030         struct cxio_rdev *rdev, *tmp;
1031
1032         t3_register_cpl_handler(CPL_ASYNC_NOTIF, NULL);
1033         list_for_each_entry_safe(rdev, tmp, &rdev_list, entry)
1034                 cxio_rdev_close(rdev);
1035         cxio_hal_destroy_rhdl_resource();
1036 }
1037
1038 static void flush_completed_wrs(struct t3_wq *wq, struct t3_cq *cq)
1039 {
1040         struct t3_swsq *sqp;
1041         __u32 ptr = wq->sq_rptr;
1042         int count = Q_COUNT(wq->sq_rptr, wq->sq_wptr);
1043
1044         sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
1045         while (count--)
1046                 if (!sqp->signaled) {
1047                         ptr++;
1048                         sqp = wq->sq + Q_PTR2IDX(ptr,  wq->sq_size_log2);
1049                 } else if (sqp->complete) {
1050
1051                         /*
1052                          * Insert this completed cqe into the swcq.
1053                          */
1054                         PDBG("%s moving cqe into swcq sq idx %ld cq idx %ld\n",
1055                              __func__, Q_PTR2IDX(ptr,  wq->sq_size_log2),
1056                              Q_PTR2IDX(cq->sw_wptr, cq->size_log2));
1057                         sqp->cqe.header |= htonl(V_CQE_SWCQE(1));
1058                         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2))
1059                                 = sqp->cqe;
1060                         cq->sw_wptr++;
1061                         sqp->signaled = 0;
1062                         break;
1063                 } else
1064                         break;
1065 }
1066
1067 static void create_read_req_cqe(struct t3_wq *wq, struct t3_cqe *hw_cqe,
1068                                 struct t3_cqe *read_cqe)
1069 {
1070         read_cqe->u.scqe.wrid_hi = wq->oldest_read->sq_wptr;
1071         read_cqe->len = wq->oldest_read->read_len;
1072         read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(*hw_cqe)) |
1073                                  V_CQE_SWCQE(SW_CQE(*hw_cqe)) |
1074                                  V_CQE_OPCODE(T3_READ_REQ) |
1075                                  V_CQE_TYPE(1));
1076 }
1077
1078 /*
1079  * Return a ptr to the next read wr in the SWSQ or NULL.
1080  */
1081 static void advance_oldest_read(struct t3_wq *wq)
1082 {
1083
1084         u32 rptr = wq->oldest_read - wq->sq + 1;
1085         u32 wptr = Q_PTR2IDX(wq->sq_wptr, wq->sq_size_log2);
1086
1087         while (Q_PTR2IDX(rptr, wq->sq_size_log2) != wptr) {
1088                 wq->oldest_read = wq->sq + Q_PTR2IDX(rptr, wq->sq_size_log2);
1089
1090                 if (wq->oldest_read->opcode == T3_READ_REQ)
1091                         return;
1092                 rptr++;
1093         }
1094         wq->oldest_read = NULL;
1095 }
1096
1097 /*
1098  * cxio_poll_cq
1099  *
1100  * Caller must:
1101  *     check the validity of the first CQE,
1102  *     supply the wq assicated with the qpid.
1103  *
1104  * credit: cq credit to return to sge.
1105  * cqe_flushed: 1 iff the CQE is flushed.
1106  * cqe: copy of the polled CQE.
1107  *
1108  * return value:
1109  *     0       CQE returned,
1110  *    -1       CQE skipped, try again.
1111  */
1112 int cxio_poll_cq(struct t3_wq *wq, struct t3_cq *cq, struct t3_cqe *cqe,
1113                      u8 *cqe_flushed, u64 *cookie, u32 *credit)
1114 {
1115         int ret = 0;
1116         struct t3_cqe *hw_cqe, read_cqe;
1117
1118         *cqe_flushed = 0;
1119         *credit = 0;
1120         hw_cqe = cxio_next_cqe(cq);
1121
1122         PDBG("%s CQE OOO %d qpid 0x%0x genbit %d type %d status 0x%0x"
1123              " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
1124              __func__, CQE_OOO(*hw_cqe), CQE_QPID(*hw_cqe),
1125              CQE_GENBIT(*hw_cqe), CQE_TYPE(*hw_cqe), CQE_STATUS(*hw_cqe),
1126              CQE_OPCODE(*hw_cqe), CQE_LEN(*hw_cqe), CQE_WRID_HI(*hw_cqe),
1127              CQE_WRID_LOW(*hw_cqe));
1128
1129         /*
1130          * skip cqe's not affiliated with a QP.
1131          */
1132         if (wq == NULL) {
1133                 ret = -1;
1134                 goto skip_cqe;
1135         }
1136
1137         /*
1138          * Gotta tweak READ completions:
1139          *      1) the cqe doesn't contain the sq_wptr from the wr.
1140          *      2) opcode not reflected from the wr.
1141          *      3) read_len not reflected from the wr.
1142          *      4) cq_type is RQ_TYPE not SQ_TYPE.
1143          */
1144         if (RQ_TYPE(*hw_cqe) && (CQE_OPCODE(*hw_cqe) == T3_READ_RESP)) {
1145
1146                 /*
1147                  * If this is an unsolicited read response, then the read
1148                  * was generated by the kernel driver as part of peer-2-peer
1149                  * connection setup.  So ignore the completion.
1150                  */
1151                 if (!wq->oldest_read) {
1152                         if (CQE_STATUS(*hw_cqe))
1153                                 wq->error = 1;
1154                         ret = -1;
1155                         goto skip_cqe;
1156                 }
1157
1158                 /*
1159                  * Don't write to the HWCQ, so create a new read req CQE
1160                  * in local memory.
1161                  */
1162                 create_read_req_cqe(wq, hw_cqe, &read_cqe);
1163                 hw_cqe = &read_cqe;
1164                 advance_oldest_read(wq);
1165         }
1166
1167         /*
1168          * T3A: Discard TERMINATE CQEs.
1169          */
1170         if (CQE_OPCODE(*hw_cqe) == T3_TERMINATE) {
1171                 ret = -1;
1172                 wq->error = 1;
1173                 goto skip_cqe;
1174         }
1175
1176         if (CQE_STATUS(*hw_cqe) || wq->error) {
1177                 *cqe_flushed = wq->error;
1178                 wq->error = 1;
1179
1180                 /*
1181                  * T3A inserts errors into the CQE.  We cannot return
1182                  * these as work completions.
1183                  */
1184                 /* incoming write failures */
1185                 if ((CQE_OPCODE(*hw_cqe) == T3_RDMA_WRITE)
1186                      && RQ_TYPE(*hw_cqe)) {
1187                         ret = -1;
1188                         goto skip_cqe;
1189                 }
1190                 /* incoming read request failures */
1191                 if ((CQE_OPCODE(*hw_cqe) == T3_READ_RESP) && SQ_TYPE(*hw_cqe)) {
1192                         ret = -1;
1193                         goto skip_cqe;
1194                 }
1195
1196                 /* incoming SEND with no receive posted failures */
1197                 if ((CQE_OPCODE(*hw_cqe) == T3_SEND) && RQ_TYPE(*hw_cqe) &&
1198                     Q_EMPTY(wq->rq_rptr, wq->rq_wptr)) {
1199                         ret = -1;
1200                         goto skip_cqe;
1201                 }
1202                 goto proc_cqe;
1203         }
1204
1205         /*
1206          * RECV completion.
1207          */
1208         if (RQ_TYPE(*hw_cqe)) {
1209
1210                 /*
1211                  * HW only validates 4 bits of MSN.  So we must validate that
1212                  * the MSN in the SEND is the next expected MSN.  If its not,
1213                  * then we complete this with TPT_ERR_MSN and mark the wq in
1214                  * error.
1215                  */
1216                 if (unlikely((CQE_WRID_MSN(*hw_cqe) != (wq->rq_rptr + 1)))) {
1217                         wq->error = 1;
1218                         hw_cqe->header |= htonl(V_CQE_STATUS(TPT_ERR_MSN));
1219                         goto proc_cqe;
1220                 }
1221                 goto proc_cqe;
1222         }
1223
1224         /*
1225          * If we get here its a send completion.
1226          *
1227          * Handle out of order completion. These get stuffed
1228          * in the SW SQ. Then the SW SQ is walked to move any
1229          * now in-order completions into the SW CQ.  This handles
1230          * 2 cases:
1231          *      1) reaping unsignaled WRs when the first subsequent
1232          *         signaled WR is completed.
1233          *      2) out of order read completions.
1234          */
1235         if (!SW_CQE(*hw_cqe) && (CQE_WRID_SQ_WPTR(*hw_cqe) != wq->sq_rptr)) {
1236                 struct t3_swsq *sqp;
1237
1238                 PDBG("%s out of order completion going in swsq at idx %ld\n",
1239                      __func__,
1240                      Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2));
1241                 sqp = wq->sq +
1242                       Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2);
1243                 sqp->cqe = *hw_cqe;
1244                 sqp->complete = 1;
1245                 ret = -1;
1246                 goto flush_wq;
1247         }
1248
1249 proc_cqe:
1250         *cqe = *hw_cqe;
1251
1252         /*
1253          * Reap the associated WR(s) that are freed up with this
1254          * completion.
1255          */
1256         if (SQ_TYPE(*hw_cqe)) {
1257                 wq->sq_rptr = CQE_WRID_SQ_WPTR(*hw_cqe);
1258                 PDBG("%s completing sq idx %ld\n", __func__,
1259                      Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2));
1260                 *cookie = (wq->sq +
1261                            Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2))->wr_id;
1262                 wq->sq_rptr++;
1263         } else {
1264                 PDBG("%s completing rq idx %ld\n", __func__,
1265                      Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
1266                 *cookie = *(wq->rq + Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
1267                 wq->rq_rptr++;
1268         }
1269
1270 flush_wq:
1271         /*
1272          * Flush any completed cqes that are now in-order.
1273          */
1274         flush_completed_wrs(wq, cq);
1275
1276 skip_cqe:
1277         if (SW_CQE(*hw_cqe)) {
1278                 PDBG("%s cq %p cqid 0x%x skip sw cqe sw_rptr 0x%x\n",
1279                      __func__, cq, cq->cqid, cq->sw_rptr);
1280                 ++cq->sw_rptr;
1281         } else {
1282                 PDBG("%s cq %p cqid 0x%x skip hw cqe rptr 0x%x\n",
1283                      __func__, cq, cq->cqid, cq->rptr);
1284                 ++cq->rptr;
1285
1286                 /*
1287                  * T3A: compute credits.
1288                  */
1289                 if (((cq->rptr - cq->wptr) > (1 << (cq->size_log2 - 1)))
1290                     || ((cq->rptr - cq->wptr) >= 128)) {
1291                         *credit = cq->rptr - cq->wptr;
1292                         cq->wptr = cq->rptr;
1293                 }
1294         }
1295         return ret;
1296 }