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 <linux/slab.h>
41#include <net/net_namespace.h>
42
43#include "cxio_resource.h"
44#include "cxio_hal.h"
45#include "cxgb3_offload.h"
46#include "sge_defs.h"
47
48static LIST_HEAD(rdev_list);
49static cxio_hal_ev_callback_func_t cxio_ev_cb = NULL;
50
51static struct cxio_rdev *cxio_hal_find_rdev_by_name(char *dev_name)
52{
53	struct cxio_rdev *rdev;
54
55	list_for_each_entry(rdev, &rdev_list, entry)
56		if (!strcmp(rdev->dev_name, dev_name))
57			return rdev;
58	return NULL;
59}
60
61static struct cxio_rdev *cxio_hal_find_rdev_by_t3cdev(struct t3cdev *tdev)
62{
63	struct cxio_rdev *rdev;
64
65	list_for_each_entry(rdev, &rdev_list, entry)
66		if (rdev->t3cdev_p == tdev)
67			return rdev;
68	return NULL;
69}
70
71int cxio_hal_cq_op(struct cxio_rdev *rdev_p, struct t3_cq *cq,
72		   enum t3_cq_opcode op, u32 credit)
73{
74	int ret;
75	struct t3_cqe *cqe;
76	u32 rptr;
77
78	struct rdma_cq_op setup;
79	setup.id = cq->cqid;
80	setup.credits = (op == CQ_CREDIT_UPDATE) ? credit : 0;
81	setup.op = op;
82	ret = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_OP, &setup);
83
84	if ((ret < 0) || (op == CQ_CREDIT_UPDATE))
85		return ret;
86
87	/*
88	 * If the rearm returned an index other than our current index,
89	 * then there might be CQE's in flight (being DMA'd).  We must wait
90	 * here for them to complete or the consumer can miss a notification.
91	 */
92	if (Q_PTR2IDX((cq->rptr), cq->size_log2) != ret) {
93		int i=0;
94
95		rptr = cq->rptr;
96
97		/*
98		 * Keep the generation correct by bumping rptr until it
99		 * matches the index returned by the rearm - 1.
100		 */
101		while (Q_PTR2IDX((rptr+1), cq->size_log2) != ret)
102			rptr++;
103
104		/*
105		 * Now rptr is the index for the (last) cqe that was
106		 * in-flight at the time the HW rearmed the CQ.  We
107		 * spin until that CQE is valid.
108		 */
109		cqe = cq->queue + Q_PTR2IDX(rptr, cq->size_log2);
110		while (!CQ_VLD_ENTRY(rptr, cq->size_log2, cqe)) {
111			udelay(1);
112			if (i++ > 1000000) {
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
125static 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
137static 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,
149		       T3_COMPLETION_FLAG | T3_NOTIFY_FLAG, 0, qpid, 7,
150		       T3_SOPEOP);
151	wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
152	sge_cmd = qpid << 8 | 3;
153	wqe->sge_cmd = cpu_to_be64(sge_cmd);
154	skb->priority = CPL_PRIORITY_CONTROL;
155	return iwch_cxgb3_ofld_send(rdev_p->t3cdev_p, skb);
156}
157
158int cxio_create_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq, int kernel)
159{
160	struct rdma_cq_setup setup;
161	int size = (1UL << (cq->size_log2)) * sizeof(struct t3_cqe);
162
163	size += 1; /* one extra page for storing cq-in-err state */
164	cq->cqid = cxio_hal_get_cqid(rdev_p->rscp);
165	if (!cq->cqid)
166		return -ENOMEM;
167	if (kernel) {
168		cq->sw_queue = kzalloc(size, GFP_KERNEL);
169		if (!cq->sw_queue)
170			return -ENOMEM;
171	}
172	cq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev), size,
173					     &(cq->dma_addr), GFP_KERNEL);
174	if (!cq->queue) {
175		kfree(cq->sw_queue);
176		return -ENOMEM;
177	}
178	dma_unmap_addr_set(cq, mapping, cq->dma_addr);
179	memset(cq->queue, 0, size);
180	setup.id = cq->cqid;
181	setup.base_addr = (u64) (cq->dma_addr);
182	setup.size = 1UL << cq->size_log2;
183	setup.credits = 65535;
184	setup.credit_thres = 1;
185	if (rdev_p->t3cdev_p->type != T3A)
186		setup.ovfl_mode = 0;
187	else
188		setup.ovfl_mode = 1;
189	return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
190}
191
192#ifdef notyet
193int cxio_resize_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
194{
195	struct rdma_cq_setup setup;
196	setup.id = cq->cqid;
197	setup.base_addr = (u64) (cq->dma_addr);
198	setup.size = 1UL << cq->size_log2;
199	setup.credits = setup.size;
200	setup.credit_thres = setup.size;	/* TBD: overflow recovery */
201	setup.ovfl_mode = 1;
202	return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
203}
204#endif
205
206static u32 get_qpid(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
207{
208	struct cxio_qpid_list *entry;
209	u32 qpid;
210	int i;
211
212	mutex_lock(&uctx->lock);
213	if (!list_empty(&uctx->qpids)) {
214		entry = list_entry(uctx->qpids.next, struct cxio_qpid_list,
215				   entry);
216		list_del(&entry->entry);
217		qpid = entry->qpid;
218		kfree(entry);
219	} else {
220		qpid = cxio_hal_get_qpid(rdev_p->rscp);
221		if (!qpid)
222			goto out;
223		for (i = qpid+1; i & rdev_p->qpmask; i++) {
224			entry = kmalloc(sizeof *entry, GFP_KERNEL);
225			if (!entry)
226				break;
227			entry->qpid = i;
228			list_add_tail(&entry->entry, &uctx->qpids);
229		}
230	}
231out:
232	mutex_unlock(&uctx->lock);
233	PDBG("%s qpid 0x%x\n", __func__, qpid);
234	return qpid;
235}
236
237static void put_qpid(struct cxio_rdev *rdev_p, u32 qpid,
238		     struct cxio_ucontext *uctx)
239{
240	struct cxio_qpid_list *entry;
241
242	entry = kmalloc(sizeof *entry, GFP_KERNEL);
243	if (!entry)
244		return;
245	PDBG("%s qpid 0x%x\n", __func__, qpid);
246	entry->qpid = qpid;
247	mutex_lock(&uctx->lock);
248	list_add_tail(&entry->entry, &uctx->qpids);
249	mutex_unlock(&uctx->lock);
250}
251
252void cxio_release_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
253{
254	struct list_head *pos, *nxt;
255	struct cxio_qpid_list *entry;
256
257	mutex_lock(&uctx->lock);
258	list_for_each_safe(pos, nxt, &uctx->qpids) {
259		entry = list_entry(pos, struct cxio_qpid_list, entry);
260		list_del_init(&entry->entry);
261		if (!(entry->qpid & rdev_p->qpmask))
262			cxio_hal_put_qpid(rdev_p->rscp, entry->qpid);
263		kfree(entry);
264	}
265	mutex_unlock(&uctx->lock);
266}
267
268void cxio_init_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
269{
270	INIT_LIST_HEAD(&uctx->qpids);
271	mutex_init(&uctx->lock);
272}
273
274int cxio_create_qp(struct cxio_rdev *rdev_p, u32 kernel_domain,
275		   struct t3_wq *wq, struct cxio_ucontext *uctx)
276{
277	int depth = 1UL << wq->size_log2;
278	int rqsize = 1UL << wq->rq_size_log2;
279
280	wq->qpid = get_qpid(rdev_p, uctx);
281	if (!wq->qpid)
282		return -ENOMEM;
283
284	wq->rq = kzalloc(depth * sizeof(struct t3_swrq), GFP_KERNEL);
285	if (!wq->rq)
286		goto err1;
287
288	wq->rq_addr = cxio_hal_rqtpool_alloc(rdev_p, rqsize);
289	if (!wq->rq_addr)
290		goto err2;
291
292	wq->sq = kzalloc(depth * sizeof(struct t3_swsq), GFP_KERNEL);
293	if (!wq->sq)
294		goto err3;
295
296	wq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev),
297					     depth * sizeof(union t3_wr),
298					     &(wq->dma_addr), GFP_KERNEL);
299	if (!wq->queue)
300		goto err4;
301
302	memset(wq->queue, 0, depth * sizeof(union t3_wr));
303	dma_unmap_addr_set(wq, mapping, wq->dma_addr);
304	wq->doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
305	if (!kernel_domain)
306		wq->udb = (u64)rdev_p->rnic_info.udbell_physbase +
307					(wq->qpid << rdev_p->qpshift);
308	wq->rdev = rdev_p;
309	PDBG("%s qpid 0x%x doorbell 0x%p udb 0x%llx\n", __func__,
310	     wq->qpid, wq->doorbell, (unsigned long long) wq->udb);
311	return 0;
312err4:
313	kfree(wq->sq);
314err3:
315	cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, rqsize);
316err2:
317	kfree(wq->rq);
318err1:
319	put_qpid(rdev_p, wq->qpid, uctx);
320	return -ENOMEM;
321}
322
323int cxio_destroy_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
324{
325	int err;
326	err = cxio_hal_clear_cq_ctx(rdev_p, cq->cqid);
327	kfree(cq->sw_queue);
328	dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
329			  (1UL << (cq->size_log2))
330			  * sizeof(struct t3_cqe), cq->queue,
331			  dma_unmap_addr(cq, mapping));
332	cxio_hal_put_cqid(rdev_p->rscp, cq->cqid);
333	return err;
334}
335
336int cxio_destroy_qp(struct cxio_rdev *rdev_p, struct t3_wq *wq,
337		    struct cxio_ucontext *uctx)
338{
339	dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
340			  (1UL << (wq->size_log2))
341			  * sizeof(union t3_wr), wq->queue,
342			  dma_unmap_addr(wq, mapping));
343	kfree(wq->sq);
344	cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, (1UL << wq->rq_size_log2));
345	kfree(wq->rq);
346	put_qpid(rdev_p, wq->qpid, uctx);
347	return 0;
348}
349
350static void insert_recv_cqe(struct t3_wq *wq, struct t3_cq *cq)
351{
352	struct t3_cqe cqe;
353
354	PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __func__,
355	     wq, cq, cq->sw_rptr, cq->sw_wptr);
356	memset(&cqe, 0, sizeof(cqe));
357	cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
358			         V_CQE_OPCODE(T3_SEND) |
359				 V_CQE_TYPE(0) |
360				 V_CQE_SWCQE(1) |
361				 V_CQE_QPID(wq->qpid) |
362				 V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
363						       cq->size_log2)));
364	*(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
365	cq->sw_wptr++;
366}
367
368int cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count)
369{
370	u32 ptr;
371	int flushed = 0;
372
373	PDBG("%s wq %p cq %p\n", __func__, wq, cq);
374
375	/* flush RQ */
376	PDBG("%s rq_rptr %u rq_wptr %u skip count %u\n", __func__,
377	    wq->rq_rptr, wq->rq_wptr, count);
378	ptr = wq->rq_rptr + count;
379	while (ptr++ != wq->rq_wptr) {
380		insert_recv_cqe(wq, cq);
381		flushed++;
382	}
383	return flushed;
384}
385
386static void insert_sq_cqe(struct t3_wq *wq, struct t3_cq *cq,
387		          struct t3_swsq *sqp)
388{
389	struct t3_cqe cqe;
390
391	PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __func__,
392	     wq, cq, cq->sw_rptr, cq->sw_wptr);
393	memset(&cqe, 0, sizeof(cqe));
394	cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
395			         V_CQE_OPCODE(sqp->opcode) |
396			         V_CQE_TYPE(1) |
397			         V_CQE_SWCQE(1) |
398			         V_CQE_QPID(wq->qpid) |
399			         V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
400						       cq->size_log2)));
401	cqe.u.scqe.wrid_hi = sqp->sq_wptr;
402
403	*(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
404	cq->sw_wptr++;
405}
406
407int cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count)
408{
409	__u32 ptr;
410	int flushed = 0;
411	struct t3_swsq *sqp = wq->sq + Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2);
412
413	ptr = wq->sq_rptr + count;
414	sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
415	while (ptr != wq->sq_wptr) {
416		sqp->signaled = 0;
417		insert_sq_cqe(wq, cq, sqp);
418		ptr++;
419		sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
420		flushed++;
421	}
422	return flushed;
423}
424
425/*
426 * Move all CQEs from the HWCQ into the SWCQ.
427 */
428void cxio_flush_hw_cq(struct t3_cq *cq)
429{
430	struct t3_cqe *cqe, *swcqe;
431
432	PDBG("%s cq %p cqid 0x%x\n", __func__, cq, cq->cqid);
433	cqe = cxio_next_hw_cqe(cq);
434	while (cqe) {
435		PDBG("%s flushing hwcq rptr 0x%x to swcq wptr 0x%x\n",
436		     __func__, cq->rptr, cq->sw_wptr);
437		swcqe = cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2);
438		*swcqe = *cqe;
439		swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
440		cq->sw_wptr++;
441		cq->rptr++;
442		cqe = cxio_next_hw_cqe(cq);
443	}
444}
445
446static int cqe_completes_wr(struct t3_cqe *cqe, struct t3_wq *wq)
447{
448	if (CQE_OPCODE(*cqe) == T3_TERMINATE)
449		return 0;
450
451	if ((CQE_OPCODE(*cqe) == T3_RDMA_WRITE) && RQ_TYPE(*cqe))
452		return 0;
453
454	if ((CQE_OPCODE(*cqe) == T3_READ_RESP) && SQ_TYPE(*cqe))
455		return 0;
456
457	if (CQE_SEND_OPCODE(*cqe) && RQ_TYPE(*cqe) &&
458	    Q_EMPTY(wq->rq_rptr, wq->rq_wptr))
459		return 0;
460
461	return 1;
462}
463
464void cxio_count_scqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
465{
466	struct t3_cqe *cqe;
467	u32 ptr;
468
469	*count = 0;
470	ptr = cq->sw_rptr;
471	while (!Q_EMPTY(ptr, cq->sw_wptr)) {
472		cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
473		if ((SQ_TYPE(*cqe) ||
474		     ((CQE_OPCODE(*cqe) == T3_READ_RESP) && wq->oldest_read)) &&
475		    (CQE_QPID(*cqe) == wq->qpid))
476			(*count)++;
477		ptr++;
478	}
479	PDBG("%s cq %p count %d\n", __func__, cq, *count);
480}
481
482void cxio_count_rcqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
483{
484	struct t3_cqe *cqe;
485	u32 ptr;
486
487	*count = 0;
488	PDBG("%s count zero %d\n", __func__, *count);
489	ptr = cq->sw_rptr;
490	while (!Q_EMPTY(ptr, cq->sw_wptr)) {
491		cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
492		if (RQ_TYPE(*cqe) && (CQE_OPCODE(*cqe) != T3_READ_RESP) &&
493		    (CQE_QPID(*cqe) == wq->qpid) && cqe_completes_wr(cqe, wq))
494			(*count)++;
495		ptr++;
496	}
497	PDBG("%s cq %p count %d\n", __func__, cq, *count);
498}
499
500static int cxio_hal_init_ctrl_cq(struct cxio_rdev *rdev_p)
501{
502	struct rdma_cq_setup setup;
503	setup.id = 0;
504	setup.base_addr = 0;	/* NULL address */
505	setup.size = 1;		/* enable the CQ */
506	setup.credits = 0;
507
508	/* force SGE to redirect to RspQ and interrupt */
509	setup.credit_thres = 0;
510	setup.ovfl_mode = 1;
511	return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
512}
513
514static int cxio_hal_init_ctrl_qp(struct cxio_rdev *rdev_p)
515{
516	int err;
517	u64 sge_cmd, ctx0, ctx1;
518	u64 base_addr;
519	struct t3_modify_qp_wr *wqe;
520	struct sk_buff *skb;
521
522	skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
523	if (!skb) {
524		PDBG("%s alloc_skb failed\n", __func__);
525		return -ENOMEM;
526	}
527	err = cxio_hal_init_ctrl_cq(rdev_p);
528	if (err) {
529		PDBG("%s err %d initializing ctrl_cq\n", __func__, err);
530		goto err;
531	}
532	rdev_p->ctrl_qp.workq = dma_alloc_coherent(
533					&(rdev_p->rnic_info.pdev->dev),
534					(1 << T3_CTRL_QP_SIZE_LOG2) *
535					sizeof(union t3_wr),
536					&(rdev_p->ctrl_qp.dma_addr),
537					GFP_KERNEL);
538	if (!rdev_p->ctrl_qp.workq) {
539		PDBG("%s dma_alloc_coherent failed\n", __func__);
540		err = -ENOMEM;
541		goto err;
542	}
543	dma_unmap_addr_set(&rdev_p->ctrl_qp, mapping,
544			   rdev_p->ctrl_qp.dma_addr);
545	rdev_p->ctrl_qp.doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
546	memset(rdev_p->ctrl_qp.workq, 0,
547	       (1 << T3_CTRL_QP_SIZE_LOG2) * sizeof(union t3_wr));
548
549	mutex_init(&rdev_p->ctrl_qp.lock);
550	init_waitqueue_head(&rdev_p->ctrl_qp.waitq);
551
552	/* update HW Ctrl QP context */
553	base_addr = rdev_p->ctrl_qp.dma_addr;
554	base_addr >>= 12;
555	ctx0 = (V_EC_SIZE((1 << T3_CTRL_QP_SIZE_LOG2)) |
556		V_EC_BASE_LO((u32) base_addr & 0xffff));
557	ctx0 <<= 32;
558	ctx0 |= V_EC_CREDITS(FW_WR_NUM);
559	base_addr >>= 16;
560	ctx1 = (u32) base_addr;
561	base_addr >>= 32;
562	ctx1 |= ((u64) (V_EC_BASE_HI((u32) base_addr & 0xf) | V_EC_RESPQ(0) |
563			V_EC_TYPE(0) | V_EC_GEN(1) |
564			V_EC_UP_TOKEN(T3_CTL_QP_TID) | F_EC_VALID)) << 32;
565	wqe = (struct t3_modify_qp_wr *) skb_put(skb, sizeof(*wqe));
566	memset(wqe, 0, sizeof(*wqe));
567	build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD, 0, 0,
568		       T3_CTL_QP_TID, 7, T3_SOPEOP);
569	wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
570	sge_cmd = (3ULL << 56) | FW_RI_SGEEC_START << 8 | 3;
571	wqe->sge_cmd = cpu_to_be64(sge_cmd);
572	wqe->ctx1 = cpu_to_be64(ctx1);
573	wqe->ctx0 = cpu_to_be64(ctx0);
574	PDBG("CtrlQP dma_addr 0x%llx workq %p size %d\n",
575	     (unsigned long long) rdev_p->ctrl_qp.dma_addr,
576	     rdev_p->ctrl_qp.workq, 1 << T3_CTRL_QP_SIZE_LOG2);
577	skb->priority = CPL_PRIORITY_CONTROL;
578	return iwch_cxgb3_ofld_send(rdev_p->t3cdev_p, skb);
579err:
580	kfree_skb(skb);
581	return err;
582}
583
584static int cxio_hal_destroy_ctrl_qp(struct cxio_rdev *rdev_p)
585{
586	dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
587			  (1UL << T3_CTRL_QP_SIZE_LOG2)
588			  * sizeof(union t3_wr), rdev_p->ctrl_qp.workq,
589			  dma_unmap_addr(&rdev_p->ctrl_qp, mapping));
590	return cxio_hal_clear_qp_ctx(rdev_p, T3_CTRL_QP_ID);
591}
592
593/* write len bytes of data into addr (32B aligned address)
594 * If data is NULL, clear len byte of memory to zero.
595 * caller acquires the ctrl_qp lock before the call
596 */
597static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev *rdev_p, u32 addr,
598				      u32 len, void *data)
599{
600	u32 i, nr_wqe, copy_len;
601	u8 *copy_data;
602	u8 wr_len, utx_len;	/* length in 8 byte flit */
603	enum t3_wr_flags flag;
604	__be64 *wqe;
605	u64 utx_cmd;
606	addr &= 0x7FFFFFF;
607	nr_wqe = len % 96 ? len / 96 + 1 : len / 96;	/* 96B max per WQE */
608	PDBG("%s wptr 0x%x rptr 0x%x len %d, nr_wqe %d data %p addr 0x%0x\n",
609	     __func__, rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, len,
610	     nr_wqe, data, addr);
611	utx_len = 3;		/* in 32B unit */
612	for (i = 0; i < nr_wqe; i++) {
613		if (Q_FULL(rdev_p->ctrl_qp.rptr, rdev_p->ctrl_qp.wptr,
614		           T3_CTRL_QP_SIZE_LOG2)) {
615			PDBG("%s ctrl_qp full wtpr 0x%0x rptr 0x%0x, "
616			     "wait for more space i %d\n", __func__,
617			     rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, i);
618			if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
619					     !Q_FULL(rdev_p->ctrl_qp.rptr,
620						     rdev_p->ctrl_qp.wptr,
621						     T3_CTRL_QP_SIZE_LOG2))) {
622				PDBG("%s ctrl_qp workq interrupted\n",
623				     __func__);
624				return -ERESTARTSYS;
625			}
626			PDBG("%s ctrl_qp wakeup, continue posting work request "
627			     "i %d\n", __func__, i);
628		}
629		wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
630						(1 << T3_CTRL_QP_SIZE_LOG2)));
631		flag = 0;
632		if (i == (nr_wqe - 1)) {
633			/* last WQE */
634			flag = T3_COMPLETION_FLAG;
635			if (len % 32)
636				utx_len = len / 32 + 1;
637			else
638				utx_len = len / 32;
639		}
640
641		/*
642		 * Force a CQE to return the credit to the workq in case
643		 * we posted more than half the max QP size of WRs
644		 */
645		if ((i != 0) &&
646		    (i % (((1 << T3_CTRL_QP_SIZE_LOG2)) >> 1) == 0)) {
647			flag = T3_COMPLETION_FLAG;
648			PDBG("%s force completion at i %d\n", __func__, i);
649		}
650
651		/* build the utx mem command */
652		wqe += (sizeof(struct t3_bypass_wr) >> 3);
653		utx_cmd = (T3_UTX_MEM_WRITE << 28) | (addr + i * 3);
654		utx_cmd <<= 32;
655		utx_cmd |= (utx_len << 28) | ((utx_len << 2) + 1);
656		*wqe = cpu_to_be64(utx_cmd);
657		wqe++;
658		copy_data = (u8 *) data + i * 96;
659		copy_len = len > 96 ? 96 : len;
660
661		/* clear memory content if data is NULL */
662		if (data)
663			memcpy(wqe, copy_data, copy_len);
664		else
665			memset(wqe, 0, copy_len);
666		if (copy_len % 32)
667			memset(((u8 *) wqe) + copy_len, 0,
668			       32 - (copy_len % 32));
669		wr_len = ((sizeof(struct t3_bypass_wr)) >> 3) + 1 +
670			 (utx_len << 2);
671		wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
672			      (1 << T3_CTRL_QP_SIZE_LOG2)));
673
674		/* wptr in the WRID[31:0] */
675		((union t3_wrid *)(wqe+1))->id0.low = rdev_p->ctrl_qp.wptr;
676
677		/*
678		 * This must be the last write with a memory barrier
679		 * for the genbit
680		 */
681		build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_BP, flag,
682			       Q_GENBIT(rdev_p->ctrl_qp.wptr,
683					T3_CTRL_QP_SIZE_LOG2), T3_CTRL_QP_ID,
684			       wr_len, T3_SOPEOP);
685		if (flag == T3_COMPLETION_FLAG)
686			ring_doorbell(rdev_p->ctrl_qp.doorbell, T3_CTRL_QP_ID);
687		len -= 96;
688		rdev_p->ctrl_qp.wptr++;
689	}
690	return 0;
691}
692
693/* IN: stag key, pdid, perm, zbva, to, len, page_size, pbl_size and pbl_addr
694 * OUT: stag index
695 * TBD: shared memory region support
696 */
697static int __cxio_tpt_op(struct cxio_rdev *rdev_p, u32 reset_tpt_entry,
698			 u32 *stag, u8 stag_state, u32 pdid,
699			 enum tpt_mem_type type, enum tpt_mem_perm perm,
700			 u32 zbva, u64 to, u32 len, u8 page_size,
701			 u32 pbl_size, u32 pbl_addr)
702{
703	int err;
704	struct tpt_entry tpt;
705	u32 stag_idx;
706	u32 wptr;
707
708	if (cxio_fatal_error(rdev_p))
709		return -EIO;
710
711	stag_state = stag_state > 0;
712	stag_idx = (*stag) >> 8;
713
714	if ((!reset_tpt_entry) && !(*stag != T3_STAG_UNSET)) {
715		stag_idx = cxio_hal_get_stag(rdev_p->rscp);
716		if (!stag_idx)
717			return -ENOMEM;
718		*stag = (stag_idx << 8) | ((*stag) & 0xFF);
719	}
720	PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
721	     __func__, stag_state, type, pdid, stag_idx);
722
723	mutex_lock(&rdev_p->ctrl_qp.lock);
724
725	/* write TPT entry */
726	if (reset_tpt_entry)
727		memset(&tpt, 0, sizeof(tpt));
728	else {
729		tpt.valid_stag_pdid = cpu_to_be32(F_TPT_VALID |
730				V_TPT_STAG_KEY((*stag) & M_TPT_STAG_KEY) |
731				V_TPT_STAG_STATE(stag_state) |
732				V_TPT_STAG_TYPE(type) | V_TPT_PDID(pdid));
733		BUG_ON(page_size >= 28);
734		tpt.flags_pagesize_qpid = cpu_to_be32(V_TPT_PERM(perm) |
735			((perm & TPT_MW_BIND) ? F_TPT_MW_BIND_ENABLE : 0) |
736			V_TPT_ADDR_TYPE((zbva ? TPT_ZBTO : TPT_VATO)) |
737			V_TPT_PAGE_SIZE(page_size));
738		tpt.rsvd_pbl_addr = cpu_to_be32(V_TPT_PBL_ADDR(PBL_OFF(rdev_p, pbl_addr)>>3));
739		tpt.len = cpu_to_be32(len);
740		tpt.va_hi = cpu_to_be32((u32) (to >> 32));
741		tpt.va_low_or_fbo = cpu_to_be32((u32) (to & 0xFFFFFFFFULL));
742		tpt.rsvd_bind_cnt_or_pstag = 0;
743		tpt.rsvd_pbl_size = cpu_to_be32(V_TPT_PBL_SIZE(pbl_size >> 2));
744	}
745	err = cxio_hal_ctrl_qp_write_mem(rdev_p,
746				       stag_idx +
747				       (rdev_p->rnic_info.tpt_base >> 5),
748				       sizeof(tpt), &tpt);
749
750	/* release the stag index to free pool */
751	if (reset_tpt_entry)
752		cxio_hal_put_stag(rdev_p->rscp, stag_idx);
753
754	wptr = rdev_p->ctrl_qp.wptr;
755	mutex_unlock(&rdev_p->ctrl_qp.lock);
756	if (!err)
757		if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
758					     SEQ32_GE(rdev_p->ctrl_qp.rptr,
759						      wptr)))
760			return -ERESTARTSYS;
761	return err;
762}
763
764int cxio_write_pbl(struct cxio_rdev *rdev_p, __be64 *pbl,
765		   u32 pbl_addr, u32 pbl_size)
766{
767	u32 wptr;
768	int err;
769
770	PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
771	     __func__, pbl_addr, rdev_p->rnic_info.pbl_base,
772	     pbl_size);
773
774	mutex_lock(&rdev_p->ctrl_qp.lock);
775	err = cxio_hal_ctrl_qp_write_mem(rdev_p, pbl_addr >> 5, pbl_size << 3,
776					 pbl);
777	wptr = rdev_p->ctrl_qp.wptr;
778	mutex_unlock(&rdev_p->ctrl_qp.lock);
779	if (err)
780		return err;
781
782	if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
783				     SEQ32_GE(rdev_p->ctrl_qp.rptr,
784					      wptr)))
785		return -ERESTARTSYS;
786
787	return 0;
788}
789
790int cxio_register_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
791			   enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
792			   u8 page_size, u32 pbl_size, u32 pbl_addr)
793{
794	*stag = T3_STAG_UNSET;
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
799int cxio_reregister_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
800			   enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
801			   u8 page_size, u32 pbl_size, u32 pbl_addr)
802{
803	return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
804			     zbva, to, len, page_size, pbl_size, pbl_addr);
805}
806
807int cxio_dereg_mem(struct cxio_rdev *rdev_p, u32 stag, u32 pbl_size,
808		   u32 pbl_addr)
809{
810	return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
811			     pbl_size, pbl_addr);
812}
813
814int cxio_allocate_window(struct cxio_rdev *rdev_p, u32 * stag, u32 pdid)
815{
816	*stag = T3_STAG_UNSET;
817	return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_MW, 0, 0, 0ULL, 0, 0,
818			     0, 0);
819}
820
821int cxio_deallocate_window(struct cxio_rdev *rdev_p, u32 stag)
822{
823	return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
824			     0, 0);
825}
826
827int cxio_allocate_stag(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid, u32 pbl_size, u32 pbl_addr)
828{
829	*stag = T3_STAG_UNSET;
830	return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_NON_SHARED_MR,
831			     0, 0, 0ULL, 0, 0, pbl_size, pbl_addr);
832}
833
834int cxio_rdma_init(struct cxio_rdev *rdev_p, struct t3_rdma_init_attr *attr)
835{
836	struct t3_rdma_init_wr *wqe;
837	struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_ATOMIC);
838	if (!skb)
839		return -ENOMEM;
840	PDBG("%s rdev_p %p\n", __func__, rdev_p);
841	wqe = (struct t3_rdma_init_wr *) __skb_put(skb, sizeof(*wqe));
842	wqe->wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_INIT));
843	wqe->wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(attr->tid) |
844					   V_FW_RIWR_LEN(sizeof(*wqe) >> 3));
845	wqe->wrid.id1 = 0;
846	wqe->qpid = cpu_to_be32(attr->qpid);
847	wqe->pdid = cpu_to_be32(attr->pdid);
848	wqe->scqid = cpu_to_be32(attr->scqid);
849	wqe->rcqid = cpu_to_be32(attr->rcqid);
850	wqe->rq_addr = cpu_to_be32(attr->rq_addr - rdev_p->rnic_info.rqt_base);
851	wqe->rq_size = cpu_to_be32(attr->rq_size);
852	wqe->mpaattrs = attr->mpaattrs;
853	wqe->qpcaps = attr->qpcaps;
854	wqe->ulpdu_size = cpu_to_be16(attr->tcp_emss);
855	wqe->rqe_count = cpu_to_be16(attr->rqe_count);
856	wqe->flags_rtr_type = cpu_to_be16(attr->flags |
857					  V_RTR_TYPE(attr->rtr_type) |
858					  V_CHAN(attr->chan));
859	wqe->ord = cpu_to_be32(attr->ord);
860	wqe->ird = cpu_to_be32(attr->ird);
861	wqe->qp_dma_addr = cpu_to_be64(attr->qp_dma_addr);
862	wqe->qp_dma_size = cpu_to_be32(attr->qp_dma_size);
863	wqe->irs = cpu_to_be32(attr->irs);
864	skb->priority = 0;	/* 0=>ToeQ; 1=>CtrlQ */
865	return iwch_cxgb3_ofld_send(rdev_p->t3cdev_p, skb);
866}
867
868void cxio_register_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
869{
870	cxio_ev_cb = ev_cb;
871}
872
873void cxio_unregister_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
874{
875	cxio_ev_cb = NULL;
876}
877
878static int cxio_hal_ev_handler(struct t3cdev *t3cdev_p, struct sk_buff *skb)
879{
880	static int cnt;
881	struct cxio_rdev *rdev_p = NULL;
882	struct respQ_msg_t *rsp_msg = (struct respQ_msg_t *) skb->data;
883	PDBG("%d: %s cq_id 0x%x cq_ptr 0x%x genbit %0x overflow %0x an %0x"
884	     " se %0x notify %0x cqbranch %0x creditth %0x\n",
885	     cnt, __func__, RSPQ_CQID(rsp_msg), RSPQ_CQPTR(rsp_msg),
886	     RSPQ_GENBIT(rsp_msg), RSPQ_OVERFLOW(rsp_msg), RSPQ_AN(rsp_msg),
887	     RSPQ_SE(rsp_msg), RSPQ_NOTIFY(rsp_msg), RSPQ_CQBRANCH(rsp_msg),
888	     RSPQ_CREDIT_THRESH(rsp_msg));
889	PDBG("CQE: QPID 0x%0x genbit %0x type 0x%0x status 0x%0x opcode %d "
890	     "len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
891	     CQE_QPID(rsp_msg->cqe), CQE_GENBIT(rsp_msg->cqe),
892	     CQE_TYPE(rsp_msg->cqe), CQE_STATUS(rsp_msg->cqe),
893	     CQE_OPCODE(rsp_msg->cqe), CQE_LEN(rsp_msg->cqe),
894	     CQE_WRID_HI(rsp_msg->cqe), CQE_WRID_LOW(rsp_msg->cqe));
895	rdev_p = (struct cxio_rdev *)t3cdev_p->ulp;
896	if (!rdev_p) {
897		PDBG("%s called by t3cdev %p with null ulp\n", __func__,
898		     t3cdev_p);
899		return 0;
900	}
901	if (CQE_QPID(rsp_msg->cqe) == T3_CTRL_QP_ID) {
902		rdev_p->ctrl_qp.rptr = CQE_WRID_LOW(rsp_msg->cqe) + 1;
903		wake_up_interruptible(&rdev_p->ctrl_qp.waitq);
904		dev_kfree_skb_irq(skb);
905	} else if (CQE_QPID(rsp_msg->cqe) == 0xfff8)
906		dev_kfree_skb_irq(skb);
907	else if (cxio_ev_cb)
908		(*cxio_ev_cb) (rdev_p, skb);
909	else
910		dev_kfree_skb_irq(skb);
911	cnt++;
912	return 0;
913}
914
915/* Caller takes care of locking if needed */
916int cxio_rdev_open(struct cxio_rdev *rdev_p)
917{
918	struct net_device *netdev_p = NULL;
919	int err = 0;
920	if (strlen(rdev_p->dev_name)) {
921		if (cxio_hal_find_rdev_by_name(rdev_p->dev_name)) {
922			return -EBUSY;
923		}
924		netdev_p = dev_get_by_name(&init_net, rdev_p->dev_name);
925		if (!netdev_p) {
926			return -EINVAL;
927		}
928		dev_put(netdev_p);
929	} else if (rdev_p->t3cdev_p) {
930		if (cxio_hal_find_rdev_by_t3cdev(rdev_p->t3cdev_p)) {
931			return -EBUSY;
932		}
933		netdev_p = rdev_p->t3cdev_p->lldev;
934		strncpy(rdev_p->dev_name, rdev_p->t3cdev_p->name,
935			T3_MAX_DEV_NAME_LEN);
936	} else {
937		PDBG("%s t3cdev_p or dev_name must be set\n", __func__);
938		return -EINVAL;
939	}
940
941	list_add_tail(&rdev_p->entry, &rdev_list);
942
943	PDBG("%s opening rnic dev %s\n", __func__, rdev_p->dev_name);
944	memset(&rdev_p->ctrl_qp, 0, sizeof(rdev_p->ctrl_qp));
945	if (!rdev_p->t3cdev_p)
946		rdev_p->t3cdev_p = dev2t3cdev(netdev_p);
947	rdev_p->t3cdev_p->ulp = (void *) rdev_p;
948
949	err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, GET_EMBEDDED_INFO,
950					 &(rdev_p->fw_info));
951	if (err) {
952		printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
953		     __func__, rdev_p->t3cdev_p, err);
954		goto err1;
955	}
956	if (G_FW_VERSION_MAJOR(rdev_p->fw_info.fw_vers) != CXIO_FW_MAJ) {
957		printk(KERN_ERR MOD "fatal firmware version mismatch: "
958		       "need version %u but adapter has version %u\n",
959		       CXIO_FW_MAJ,
960		       G_FW_VERSION_MAJOR(rdev_p->fw_info.fw_vers));
961		err = -EINVAL;
962		goto err1;
963	}
964
965	err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_GET_PARAMS,
966					 &(rdev_p->rnic_info));
967	if (err) {
968		printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
969		     __func__, rdev_p->t3cdev_p, err);
970		goto err1;
971	}
972	err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, GET_PORTS,
973				    &(rdev_p->port_info));
974	if (err) {
975		printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
976		     __func__, rdev_p->t3cdev_p, err);
977		goto err1;
978	}
979
980	/*
981	 * qpshift is the number of bits to shift the qpid left in order
982	 * to get the correct address of the doorbell for that qp.
983	 */
984	cxio_init_ucontext(rdev_p, &rdev_p->uctx);
985	rdev_p->qpshift = PAGE_SHIFT -
986			  ilog2(65536 >>
987			            ilog2(rdev_p->rnic_info.udbell_len >>
988					      PAGE_SHIFT));
989	rdev_p->qpnr = rdev_p->rnic_info.udbell_len >> PAGE_SHIFT;
990	rdev_p->qpmask = (65536 >> ilog2(rdev_p->qpnr)) - 1;
991	PDBG("%s rnic %s info: tpt_base 0x%0x tpt_top 0x%0x num stags %d "
992	     "pbl_base 0x%0x pbl_top 0x%0x rqt_base 0x%0x, rqt_top 0x%0x\n",
993	     __func__, rdev_p->dev_name, rdev_p->rnic_info.tpt_base,
994	     rdev_p->rnic_info.tpt_top, cxio_num_stags(rdev_p),
995	     rdev_p->rnic_info.pbl_base,
996	     rdev_p->rnic_info.pbl_top, rdev_p->rnic_info.rqt_base,
997	     rdev_p->rnic_info.rqt_top);
998	PDBG("udbell_len 0x%0x udbell_physbase 0x%lx kdb_addr %p qpshift %lu "
999	     "qpnr %d qpmask 0x%x\n",
1000	     rdev_p->rnic_info.udbell_len,
1001	     rdev_p->rnic_info.udbell_physbase, rdev_p->rnic_info.kdb_addr,
1002	     rdev_p->qpshift, rdev_p->qpnr, rdev_p->qpmask);
1003
1004	err = cxio_hal_init_ctrl_qp(rdev_p);
1005	if (err) {
1006		printk(KERN_ERR "%s error %d initializing ctrl_qp.\n",
1007		       __func__, err);
1008		goto err1;
1009	}
1010	err = cxio_hal_init_resource(rdev_p, cxio_num_stags(rdev_p), 0,
1011				     0, T3_MAX_NUM_QP, T3_MAX_NUM_CQ,
1012				     T3_MAX_NUM_PD);
1013	if (err) {
1014		printk(KERN_ERR "%s error %d initializing hal resources.\n",
1015		       __func__, err);
1016		goto err2;
1017	}
1018	err = cxio_hal_pblpool_create(rdev_p);
1019	if (err) {
1020		printk(KERN_ERR "%s error %d initializing pbl mem pool.\n",
1021		       __func__, err);
1022		goto err3;
1023	}
1024	err = cxio_hal_rqtpool_create(rdev_p);
1025	if (err) {
1026		printk(KERN_ERR "%s error %d initializing rqt mem pool.\n",
1027		       __func__, err);
1028		goto err4;
1029	}
1030	return 0;
1031err4:
1032	cxio_hal_pblpool_destroy(rdev_p);
1033err3:
1034	cxio_hal_destroy_resource(rdev_p->rscp);
1035err2:
1036	cxio_hal_destroy_ctrl_qp(rdev_p);
1037err1:
1038	rdev_p->t3cdev_p->ulp = NULL;
1039	list_del(&rdev_p->entry);
1040	return err;
1041}
1042
1043void cxio_rdev_close(struct cxio_rdev *rdev_p)
1044{
1045	if (rdev_p) {
1046		cxio_hal_pblpool_destroy(rdev_p);
1047		cxio_hal_rqtpool_destroy(rdev_p);
1048		list_del(&rdev_p->entry);
1049		cxio_hal_destroy_ctrl_qp(rdev_p);
1050		cxio_hal_destroy_resource(rdev_p->rscp);
1051		rdev_p->t3cdev_p->ulp = NULL;
1052	}
1053}
1054
1055int __init cxio_hal_init(void)
1056{
1057	if (cxio_hal_init_rhdl_resource(T3_MAX_NUM_RI))
1058		return -ENOMEM;
1059	t3_register_cpl_handler(CPL_ASYNC_NOTIF, cxio_hal_ev_handler);
1060	return 0;
1061}
1062
1063void __exit cxio_hal_exit(void)
1064{
1065	struct cxio_rdev *rdev, *tmp;
1066
1067	t3_register_cpl_handler(CPL_ASYNC_NOTIF, NULL);
1068	list_for_each_entry_safe(rdev, tmp, &rdev_list, entry)
1069		cxio_rdev_close(rdev);
1070	cxio_hal_destroy_rhdl_resource();
1071}
1072
1073static void flush_completed_wrs(struct t3_wq *wq, struct t3_cq *cq)
1074{
1075	struct t3_swsq *sqp;
1076	__u32 ptr = wq->sq_rptr;
1077	int count = Q_COUNT(wq->sq_rptr, wq->sq_wptr);
1078
1079	sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
1080	while (count--)
1081		if (!sqp->signaled) {
1082			ptr++;
1083			sqp = wq->sq + Q_PTR2IDX(ptr,  wq->sq_size_log2);
1084		} else if (sqp->complete) {
1085
1086			/*
1087			 * Insert this completed cqe into the swcq.
1088			 */
1089			PDBG("%s moving cqe into swcq sq idx %ld cq idx %ld\n",
1090			     __func__, Q_PTR2IDX(ptr,  wq->sq_size_log2),
1091			     Q_PTR2IDX(cq->sw_wptr, cq->size_log2));
1092			sqp->cqe.header |= htonl(V_CQE_SWCQE(1));
1093			*(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2))
1094				= sqp->cqe;
1095			cq->sw_wptr++;
1096			sqp->signaled = 0;
1097			break;
1098		} else
1099			break;
1100}
1101
1102static void create_read_req_cqe(struct t3_wq *wq, struct t3_cqe *hw_cqe,
1103				struct t3_cqe *read_cqe)
1104{
1105	read_cqe->u.scqe.wrid_hi = wq->oldest_read->sq_wptr;
1106	read_cqe->len = wq->oldest_read->read_len;
1107	read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(*hw_cqe)) |
1108				 V_CQE_SWCQE(SW_CQE(*hw_cqe)) |
1109				 V_CQE_OPCODE(T3_READ_REQ) |
1110				 V_CQE_TYPE(1));
1111}
1112
1113/*
1114 * Return a ptr to the next read wr in the SWSQ or NULL.
1115 */
1116static void advance_oldest_read(struct t3_wq *wq)
1117{
1118
1119	u32 rptr = wq->oldest_read - wq->sq + 1;
1120	u32 wptr = Q_PTR2IDX(wq->sq_wptr, wq->sq_size_log2);
1121
1122	while (Q_PTR2IDX(rptr, wq->sq_size_log2) != wptr) {
1123		wq->oldest_read = wq->sq + Q_PTR2IDX(rptr, wq->sq_size_log2);
1124
1125		if (wq->oldest_read->opcode == T3_READ_REQ)
1126			return;
1127		rptr++;
1128	}
1129	wq->oldest_read = NULL;
1130}
1131
1132/*
1133 * cxio_poll_cq
1134 *
1135 * Caller must:
1136 *     check the validity of the first CQE,
1137 *     supply the wq assicated with the qpid.
1138 *
1139 * credit: cq credit to return to sge.
1140 * cqe_flushed: 1 iff the CQE is flushed.
1141 * cqe: copy of the polled CQE.
1142 *
1143 * return value:
1144 *     0       CQE returned,
1145 *    -1       CQE skipped, try again.
1146 */
1147int cxio_poll_cq(struct t3_wq *wq, struct t3_cq *cq, struct t3_cqe *cqe,
1148		     u8 *cqe_flushed, u64 *cookie, u32 *credit)
1149{
1150	int ret = 0;
1151	struct t3_cqe *hw_cqe, read_cqe;
1152
1153	*cqe_flushed = 0;
1154	*credit = 0;
1155	hw_cqe = cxio_next_cqe(cq);
1156
1157	PDBG("%s CQE OOO %d qpid 0x%0x genbit %d type %d status 0x%0x"
1158	     " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
1159	     __func__, CQE_OOO(*hw_cqe), CQE_QPID(*hw_cqe),
1160	     CQE_GENBIT(*hw_cqe), CQE_TYPE(*hw_cqe), CQE_STATUS(*hw_cqe),
1161	     CQE_OPCODE(*hw_cqe), CQE_LEN(*hw_cqe), CQE_WRID_HI(*hw_cqe),
1162	     CQE_WRID_LOW(*hw_cqe));
1163
1164	/*
1165	 * skip cqe's not affiliated with a QP.
1166	 */
1167	if (wq == NULL) {
1168		ret = -1;
1169		goto skip_cqe;
1170	}
1171
1172	/*
1173	 * Gotta tweak READ completions:
1174	 *	1) the cqe doesn't contain the sq_wptr from the wr.
1175	 *	2) opcode not reflected from the wr.
1176	 *	3) read_len not reflected from the wr.
1177	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
1178	 */
1179	if (RQ_TYPE(*hw_cqe) && (CQE_OPCODE(*hw_cqe) == T3_READ_RESP)) {
1180
1181		/*
1182		 * If this is an unsolicited read response, then the read
1183		 * was generated by the kernel driver as part of peer-2-peer
1184		 * connection setup.  So ignore the completion.
1185		 */
1186		if (!wq->oldest_read) {
1187			if (CQE_STATUS(*hw_cqe))
1188				wq->error = 1;
1189			ret = -1;
1190			goto skip_cqe;
1191		}
1192
1193		/*
1194		 * Don't write to the HWCQ, so create a new read req CQE
1195		 * in local memory.
1196		 */
1197		create_read_req_cqe(wq, hw_cqe, &read_cqe);
1198		hw_cqe = &read_cqe;
1199		advance_oldest_read(wq);
1200	}
1201
1202	/*
1203	 * T3A: Discard TERMINATE CQEs.
1204	 */
1205	if (CQE_OPCODE(*hw_cqe) == T3_TERMINATE) {
1206		ret = -1;
1207		wq->error = 1;
1208		goto skip_cqe;
1209	}
1210
1211	if (CQE_STATUS(*hw_cqe) || wq->error) {
1212		*cqe_flushed = wq->error;
1213		wq->error = 1;
1214
1215		/*
1216		 * T3A inserts errors into the CQE.  We cannot return
1217		 * these as work completions.
1218		 */
1219		/* incoming write failures */
1220		if ((CQE_OPCODE(*hw_cqe) == T3_RDMA_WRITE)
1221		     && RQ_TYPE(*hw_cqe)) {
1222			ret = -1;
1223			goto skip_cqe;
1224		}
1225		/* incoming read request failures */
1226		if ((CQE_OPCODE(*hw_cqe) == T3_READ_RESP) && SQ_TYPE(*hw_cqe)) {
1227			ret = -1;
1228			goto skip_cqe;
1229		}
1230
1231		/* incoming SEND with no receive posted failures */
1232		if (CQE_SEND_OPCODE(*hw_cqe) && RQ_TYPE(*hw_cqe) &&
1233		    Q_EMPTY(wq->rq_rptr, wq->rq_wptr)) {
1234			ret = -1;
1235			goto skip_cqe;
1236		}
1237		BUG_ON((*cqe_flushed == 0) && !SW_CQE(*hw_cqe));
1238		goto proc_cqe;
1239	}
1240
1241	/*
1242	 * RECV completion.
1243	 */
1244	if (RQ_TYPE(*hw_cqe)) {
1245
1246		/*
1247		 * HW only validates 4 bits of MSN.  So we must validate that
1248		 * the MSN in the SEND is the next expected MSN.  If its not,
1249		 * then we complete this with TPT_ERR_MSN and mark the wq in
1250		 * error.
1251		 */
1252
1253		if (Q_EMPTY(wq->rq_rptr, wq->rq_wptr)) {
1254			wq->error = 1;
1255			ret = -1;
1256			goto skip_cqe;
1257		}
1258
1259		if (unlikely((CQE_WRID_MSN(*hw_cqe) != (wq->rq_rptr + 1)))) {
1260			wq->error = 1;
1261			hw_cqe->header |= htonl(V_CQE_STATUS(TPT_ERR_MSN));
1262			goto proc_cqe;
1263		}
1264		goto proc_cqe;
1265	}
1266
1267	/*
1268	 * If we get here its a send completion.
1269	 *
1270	 * Handle out of order completion. These get stuffed
1271	 * in the SW SQ. Then the SW SQ is walked to move any
1272	 * now in-order completions into the SW CQ.  This handles
1273	 * 2 cases:
1274	 *	1) reaping unsignaled WRs when the first subsequent
1275	 *	   signaled WR is completed.
1276	 *	2) out of order read completions.
1277	 */
1278	if (!SW_CQE(*hw_cqe) && (CQE_WRID_SQ_WPTR(*hw_cqe) != wq->sq_rptr)) {
1279		struct t3_swsq *sqp;
1280
1281		PDBG("%s out of order completion going in swsq at idx %ld\n",
1282		     __func__,
1283		     Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2));
1284		sqp = wq->sq +
1285		      Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2);
1286		sqp->cqe = *hw_cqe;
1287		sqp->complete = 1;
1288		ret = -1;
1289		goto flush_wq;
1290	}
1291
1292proc_cqe:
1293	*cqe = *hw_cqe;
1294
1295	/*
1296	 * Reap the associated WR(s) that are freed up with this
1297	 * completion.
1298	 */
1299	if (SQ_TYPE(*hw_cqe)) {
1300		wq->sq_rptr = CQE_WRID_SQ_WPTR(*hw_cqe);
1301		PDBG("%s completing sq idx %ld\n", __func__,
1302		     Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2));
1303		*cookie = wq->sq[Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2)].wr_id;
1304		wq->sq_rptr++;
1305	} else {
1306		PDBG("%s completing rq idx %ld\n", __func__,
1307		     Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
1308		*cookie = wq->rq[Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2)].wr_id;
1309		if (wq->rq[Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2)].pbl_addr)
1310			cxio_hal_pblpool_free(wq->rdev,
1311				wq->rq[Q_PTR2IDX(wq->rq_rptr,
1312				wq->rq_size_log2)].pbl_addr, T3_STAG0_PBL_SIZE);
1313		BUG_ON(Q_EMPTY(wq->rq_rptr, wq->rq_wptr));
1314		wq->rq_rptr++;
1315	}
1316
1317flush_wq:
1318	/*
1319	 * Flush any completed cqes that are now in-order.
1320	 */
1321	flush_completed_wrs(wq, cq);
1322
1323skip_cqe:
1324	if (SW_CQE(*hw_cqe)) {
1325		PDBG("%s cq %p cqid 0x%x skip sw cqe sw_rptr 0x%x\n",
1326		     __func__, cq, cq->cqid, cq->sw_rptr);
1327		++cq->sw_rptr;
1328	} else {
1329		PDBG("%s cq %p cqid 0x%x skip hw cqe rptr 0x%x\n",
1330		     __func__, cq, cq->cqid, cq->rptr);
1331		++cq->rptr;
1332
1333		/*
1334		 * T3A: compute credits.
1335		 */
1336		if (((cq->rptr - cq->wptr) > (1 << (cq->size_log2 - 1)))
1337		    || ((cq->rptr - cq->wptr) >= 128)) {
1338			*credit = cq->rptr - cq->wptr;
1339			cq->wptr = cq->rptr;
1340		}
1341	}
1342	return ret;
1343}
1344