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
48 static LIST_HEAD(rdev_list);
49 static cxio_hal_ev_callback_func_t cxio_ev_cb = NULL;
50
cxio_hal_find_rdev_by_name(char * dev_name)51 static 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
cxio_hal_find_rdev_by_t3cdev(struct t3cdev * tdev)61 static 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
cxio_hal_cq_op(struct cxio_rdev * rdev_p,struct t3_cq * cq,enum t3_cq_opcode op,u32 credit)71 int 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 pr_err("%s: stalled rnic\n", rdev_p->dev_name);
114 return -EIO;
115 }
116 }
117
118 return 1;
119 }
120
121 return 0;
122 }
123
cxio_hal_clear_cq_ctx(struct cxio_rdev * rdev_p,u32 cqid)124 static int cxio_hal_clear_cq_ctx(struct cxio_rdev *rdev_p, u32 cqid)
125 {
126 struct rdma_cq_setup setup;
127 setup.id = cqid;
128 setup.base_addr = 0; /* NULL address */
129 setup.size = 0; /* disaable the CQ */
130 setup.credits = 0;
131 setup.credit_thres = 0;
132 setup.ovfl_mode = 0;
133 return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
134 }
135
cxio_hal_clear_qp_ctx(struct cxio_rdev * rdev_p,u32 qpid)136 static int cxio_hal_clear_qp_ctx(struct cxio_rdev *rdev_p, u32 qpid)
137 {
138 u64 sge_cmd;
139 struct t3_modify_qp_wr *wqe;
140 struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
141 if (!skb) {
142 pr_debug("%s alloc_skb failed\n", __func__);
143 return -ENOMEM;
144 }
145 wqe = skb_put_zero(skb, sizeof(*wqe));
146 build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD,
147 T3_COMPLETION_FLAG | T3_NOTIFY_FLAG, 0, qpid, 7,
148 T3_SOPEOP);
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 iwch_cxgb3_ofld_send(rdev_p->t3cdev_p, skb);
154 }
155
cxio_create_cq(struct cxio_rdev * rdev_p,struct t3_cq * cq,int kernel)156 int cxio_create_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq, int kernel)
157 {
158 struct rdma_cq_setup setup;
159 int size = (1UL << (cq->size_log2)) * sizeof(struct t3_cqe);
160
161 size += 1; /* one extra page for storing cq-in-err state */
162 cq->cqid = cxio_hal_get_cqid(rdev_p->rscp);
163 if (!cq->cqid)
164 return -ENOMEM;
165 if (kernel) {
166 cq->sw_queue = kzalloc(size, GFP_KERNEL);
167 if (!cq->sw_queue)
168 return -ENOMEM;
169 }
170 cq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev), size,
171 &(cq->dma_addr), GFP_KERNEL);
172 if (!cq->queue) {
173 kfree(cq->sw_queue);
174 return -ENOMEM;
175 }
176 dma_unmap_addr_set(cq, mapping, cq->dma_addr);
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
get_qpid(struct cxio_rdev * rdev_p,struct cxio_ucontext * uctx)189 static u32 get_qpid(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
190 {
191 struct cxio_qpid_list *entry;
192 u32 qpid;
193 int i;
194
195 mutex_lock(&uctx->lock);
196 if (!list_empty(&uctx->qpids)) {
197 entry = list_entry(uctx->qpids.next, struct cxio_qpid_list,
198 entry);
199 list_del(&entry->entry);
200 qpid = entry->qpid;
201 kfree(entry);
202 } else {
203 qpid = cxio_hal_get_qpid(rdev_p->rscp);
204 if (!qpid)
205 goto out;
206 for (i = qpid+1; i & rdev_p->qpmask; i++) {
207 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
208 if (!entry)
209 break;
210 entry->qpid = i;
211 list_add_tail(&entry->entry, &uctx->qpids);
212 }
213 }
214 out:
215 mutex_unlock(&uctx->lock);
216 pr_debug("%s qpid 0x%x\n", __func__, qpid);
217 return qpid;
218 }
219
put_qpid(struct cxio_rdev * rdev_p,u32 qpid,struct cxio_ucontext * uctx)220 static void put_qpid(struct cxio_rdev *rdev_p, u32 qpid,
221 struct cxio_ucontext *uctx)
222 {
223 struct cxio_qpid_list *entry;
224
225 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
226 if (!entry)
227 return;
228 pr_debug("%s qpid 0x%x\n", __func__, qpid);
229 entry->qpid = qpid;
230 mutex_lock(&uctx->lock);
231 list_add_tail(&entry->entry, &uctx->qpids);
232 mutex_unlock(&uctx->lock);
233 }
234
cxio_release_ucontext(struct cxio_rdev * rdev_p,struct cxio_ucontext * uctx)235 void cxio_release_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
236 {
237 struct list_head *pos, *nxt;
238 struct cxio_qpid_list *entry;
239
240 mutex_lock(&uctx->lock);
241 list_for_each_safe(pos, nxt, &uctx->qpids) {
242 entry = list_entry(pos, struct cxio_qpid_list, entry);
243 list_del_init(&entry->entry);
244 if (!(entry->qpid & rdev_p->qpmask))
245 cxio_hal_put_qpid(rdev_p->rscp, entry->qpid);
246 kfree(entry);
247 }
248 mutex_unlock(&uctx->lock);
249 }
250
cxio_init_ucontext(struct cxio_rdev * rdev_p,struct cxio_ucontext * uctx)251 void cxio_init_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
252 {
253 INIT_LIST_HEAD(&uctx->qpids);
254 mutex_init(&uctx->lock);
255 }
256
cxio_create_qp(struct cxio_rdev * rdev_p,u32 kernel_domain,struct t3_wq * wq,struct cxio_ucontext * uctx)257 int cxio_create_qp(struct cxio_rdev *rdev_p, u32 kernel_domain,
258 struct t3_wq *wq, struct cxio_ucontext *uctx)
259 {
260 int depth = 1UL << wq->size_log2;
261 int rqsize = 1UL << wq->rq_size_log2;
262
263 wq->qpid = get_qpid(rdev_p, uctx);
264 if (!wq->qpid)
265 return -ENOMEM;
266
267 wq->rq = kcalloc(depth, sizeof(struct t3_swrq), GFP_KERNEL);
268 if (!wq->rq)
269 goto err1;
270
271 wq->rq_addr = cxio_hal_rqtpool_alloc(rdev_p, rqsize);
272 if (!wq->rq_addr)
273 goto err2;
274
275 wq->sq = kcalloc(depth, sizeof(struct t3_swsq), GFP_KERNEL);
276 if (!wq->sq)
277 goto err3;
278
279 wq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev),
280 depth * sizeof(union t3_wr),
281 &(wq->dma_addr), GFP_KERNEL);
282 if (!wq->queue)
283 goto err4;
284
285 dma_unmap_addr_set(wq, mapping, wq->dma_addr);
286 wq->doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
287 if (!kernel_domain)
288 wq->udb = (u64)rdev_p->rnic_info.udbell_physbase +
289 (wq->qpid << rdev_p->qpshift);
290 wq->rdev = rdev_p;
291 pr_debug("%s qpid 0x%x doorbell 0x%p udb 0x%llx\n",
292 __func__, wq->qpid, wq->doorbell, (unsigned long long)wq->udb);
293 return 0;
294 err4:
295 kfree(wq->sq);
296 err3:
297 cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, rqsize);
298 err2:
299 kfree(wq->rq);
300 err1:
301 put_qpid(rdev_p, wq->qpid, uctx);
302 return -ENOMEM;
303 }
304
cxio_destroy_cq(struct cxio_rdev * rdev_p,struct t3_cq * cq)305 void cxio_destroy_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
306 {
307 cxio_hal_clear_cq_ctx(rdev_p, cq->cqid);
308 kfree(cq->sw_queue);
309 dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
310 (1UL << (cq->size_log2))
311 * sizeof(struct t3_cqe) + 1, cq->queue,
312 dma_unmap_addr(cq, mapping));
313 cxio_hal_put_cqid(rdev_p->rscp, cq->cqid);
314 }
315
cxio_destroy_qp(struct cxio_rdev * rdev_p,struct t3_wq * wq,struct cxio_ucontext * uctx)316 int cxio_destroy_qp(struct cxio_rdev *rdev_p, struct t3_wq *wq,
317 struct cxio_ucontext *uctx)
318 {
319 dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
320 (1UL << (wq->size_log2))
321 * sizeof(union t3_wr), wq->queue,
322 dma_unmap_addr(wq, mapping));
323 kfree(wq->sq);
324 cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, (1UL << wq->rq_size_log2));
325 kfree(wq->rq);
326 put_qpid(rdev_p, wq->qpid, uctx);
327 return 0;
328 }
329
insert_recv_cqe(struct t3_wq * wq,struct t3_cq * cq)330 static void insert_recv_cqe(struct t3_wq *wq, struct t3_cq *cq)
331 {
332 struct t3_cqe cqe;
333
334 pr_debug("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __func__,
335 wq, cq, cq->sw_rptr, cq->sw_wptr);
336 memset(&cqe, 0, sizeof(cqe));
337 cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
338 V_CQE_OPCODE(T3_SEND) |
339 V_CQE_TYPE(0) |
340 V_CQE_SWCQE(1) |
341 V_CQE_QPID(wq->qpid) |
342 V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
343 cq->size_log2)));
344 *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
345 cq->sw_wptr++;
346 }
347
cxio_flush_rq(struct t3_wq * wq,struct t3_cq * cq,int count)348 int cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count)
349 {
350 u32 ptr;
351 int flushed = 0;
352
353 pr_debug("%s wq %p cq %p\n", __func__, wq, cq);
354
355 /* flush RQ */
356 pr_debug("%s rq_rptr %u rq_wptr %u skip count %u\n", __func__,
357 wq->rq_rptr, wq->rq_wptr, count);
358 ptr = wq->rq_rptr + count;
359 while (ptr++ != wq->rq_wptr) {
360 insert_recv_cqe(wq, cq);
361 flushed++;
362 }
363 return flushed;
364 }
365
insert_sq_cqe(struct t3_wq * wq,struct t3_cq * cq,struct t3_swsq * sqp)366 static void insert_sq_cqe(struct t3_wq *wq, struct t3_cq *cq,
367 struct t3_swsq *sqp)
368 {
369 struct t3_cqe cqe;
370
371 pr_debug("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __func__,
372 wq, cq, cq->sw_rptr, cq->sw_wptr);
373 memset(&cqe, 0, sizeof(cqe));
374 cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
375 V_CQE_OPCODE(sqp->opcode) |
376 V_CQE_TYPE(1) |
377 V_CQE_SWCQE(1) |
378 V_CQE_QPID(wq->qpid) |
379 V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
380 cq->size_log2)));
381 cqe.u.scqe.wrid_hi = sqp->sq_wptr;
382
383 *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
384 cq->sw_wptr++;
385 }
386
cxio_flush_sq(struct t3_wq * wq,struct t3_cq * cq,int count)387 int cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count)
388 {
389 __u32 ptr = wq->sq_rptr + count;
390 int flushed = 0;
391 struct t3_swsq *sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
392
393 while (ptr != wq->sq_wptr) {
394 sqp->signaled = 0;
395 insert_sq_cqe(wq, cq, sqp);
396 ptr++;
397 sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
398 flushed++;
399 }
400 return flushed;
401 }
402
403 /*
404 * Move all CQEs from the HWCQ into the SWCQ.
405 */
cxio_flush_hw_cq(struct t3_cq * cq)406 void cxio_flush_hw_cq(struct t3_cq *cq)
407 {
408 struct t3_cqe *cqe, *swcqe;
409
410 pr_debug("%s cq %p cqid 0x%x\n", __func__, cq, cq->cqid);
411 cqe = cxio_next_hw_cqe(cq);
412 while (cqe) {
413 pr_debug("%s flushing hwcq rptr 0x%x to swcq wptr 0x%x\n",
414 __func__, cq->rptr, cq->sw_wptr);
415 swcqe = cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2);
416 *swcqe = *cqe;
417 swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
418 cq->sw_wptr++;
419 cq->rptr++;
420 cqe = cxio_next_hw_cqe(cq);
421 }
422 }
423
cqe_completes_wr(struct t3_cqe * cqe,struct t3_wq * wq)424 static int cqe_completes_wr(struct t3_cqe *cqe, struct t3_wq *wq)
425 {
426 if (CQE_OPCODE(*cqe) == T3_TERMINATE)
427 return 0;
428
429 if ((CQE_OPCODE(*cqe) == T3_RDMA_WRITE) && RQ_TYPE(*cqe))
430 return 0;
431
432 if ((CQE_OPCODE(*cqe) == T3_READ_RESP) && SQ_TYPE(*cqe))
433 return 0;
434
435 if (CQE_SEND_OPCODE(*cqe) && RQ_TYPE(*cqe) &&
436 Q_EMPTY(wq->rq_rptr, wq->rq_wptr))
437 return 0;
438
439 return 1;
440 }
441
cxio_count_scqes(struct t3_cq * cq,struct t3_wq * wq,int * count)442 void cxio_count_scqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
443 {
444 struct t3_cqe *cqe;
445 u32 ptr;
446
447 *count = 0;
448 ptr = cq->sw_rptr;
449 while (!Q_EMPTY(ptr, cq->sw_wptr)) {
450 cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
451 if ((SQ_TYPE(*cqe) ||
452 ((CQE_OPCODE(*cqe) == T3_READ_RESP) && wq->oldest_read)) &&
453 (CQE_QPID(*cqe) == wq->qpid))
454 (*count)++;
455 ptr++;
456 }
457 pr_debug("%s cq %p count %d\n", __func__, cq, *count);
458 }
459
cxio_count_rcqes(struct t3_cq * cq,struct t3_wq * wq,int * count)460 void cxio_count_rcqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
461 {
462 struct t3_cqe *cqe;
463 u32 ptr;
464
465 *count = 0;
466 pr_debug("%s count zero %d\n", __func__, *count);
467 ptr = cq->sw_rptr;
468 while (!Q_EMPTY(ptr, cq->sw_wptr)) {
469 cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
470 if (RQ_TYPE(*cqe) && (CQE_OPCODE(*cqe) != T3_READ_RESP) &&
471 (CQE_QPID(*cqe) == wq->qpid) && cqe_completes_wr(cqe, wq))
472 (*count)++;
473 ptr++;
474 }
475 pr_debug("%s cq %p count %d\n", __func__, cq, *count);
476 }
477
cxio_hal_init_ctrl_cq(struct cxio_rdev * rdev_p)478 static int cxio_hal_init_ctrl_cq(struct cxio_rdev *rdev_p)
479 {
480 struct rdma_cq_setup setup;
481 setup.id = 0;
482 setup.base_addr = 0; /* NULL address */
483 setup.size = 1; /* enable the CQ */
484 setup.credits = 0;
485
486 /* force SGE to redirect to RspQ and interrupt */
487 setup.credit_thres = 0;
488 setup.ovfl_mode = 1;
489 return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
490 }
491
cxio_hal_init_ctrl_qp(struct cxio_rdev * rdev_p)492 static int cxio_hal_init_ctrl_qp(struct cxio_rdev *rdev_p)
493 {
494 int err;
495 u64 sge_cmd, ctx0, ctx1;
496 u64 base_addr;
497 struct t3_modify_qp_wr *wqe;
498 struct sk_buff *skb;
499
500 skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
501 if (!skb) {
502 pr_debug("%s alloc_skb failed\n", __func__);
503 return -ENOMEM;
504 }
505 err = cxio_hal_init_ctrl_cq(rdev_p);
506 if (err) {
507 pr_debug("%s err %d initializing ctrl_cq\n", __func__, err);
508 goto err;
509 }
510 rdev_p->ctrl_qp.workq = dma_alloc_coherent(
511 &(rdev_p->rnic_info.pdev->dev),
512 (1 << T3_CTRL_QP_SIZE_LOG2) *
513 sizeof(union t3_wr),
514 &(rdev_p->ctrl_qp.dma_addr),
515 GFP_KERNEL);
516 if (!rdev_p->ctrl_qp.workq) {
517 pr_debug("%s dma_alloc_coherent failed\n", __func__);
518 err = -ENOMEM;
519 goto err;
520 }
521 dma_unmap_addr_set(&rdev_p->ctrl_qp, mapping,
522 rdev_p->ctrl_qp.dma_addr);
523 rdev_p->ctrl_qp.doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
524
525 mutex_init(&rdev_p->ctrl_qp.lock);
526 init_waitqueue_head(&rdev_p->ctrl_qp.waitq);
527
528 /* update HW Ctrl QP context */
529 base_addr = rdev_p->ctrl_qp.dma_addr;
530 base_addr >>= 12;
531 ctx0 = (V_EC_SIZE((1 << T3_CTRL_QP_SIZE_LOG2)) |
532 V_EC_BASE_LO((u32) base_addr & 0xffff));
533 ctx0 <<= 32;
534 ctx0 |= V_EC_CREDITS(FW_WR_NUM);
535 base_addr >>= 16;
536 ctx1 = (u32) base_addr;
537 base_addr >>= 32;
538 ctx1 |= ((u64) (V_EC_BASE_HI((u32) base_addr & 0xf) | V_EC_RESPQ(0) |
539 V_EC_TYPE(0) | V_EC_GEN(1) |
540 V_EC_UP_TOKEN(T3_CTL_QP_TID) | F_EC_VALID)) << 32;
541 wqe = skb_put_zero(skb, sizeof(*wqe));
542 build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD, 0, 0,
543 T3_CTL_QP_TID, 7, T3_SOPEOP);
544 wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
545 sge_cmd = (3ULL << 56) | FW_RI_SGEEC_START << 8 | 3;
546 wqe->sge_cmd = cpu_to_be64(sge_cmd);
547 wqe->ctx1 = cpu_to_be64(ctx1);
548 wqe->ctx0 = cpu_to_be64(ctx0);
549 pr_debug("CtrlQP dma_addr %pad workq %p size %d\n",
550 &rdev_p->ctrl_qp.dma_addr, rdev_p->ctrl_qp.workq,
551 1 << T3_CTRL_QP_SIZE_LOG2);
552 skb->priority = CPL_PRIORITY_CONTROL;
553 return iwch_cxgb3_ofld_send(rdev_p->t3cdev_p, skb);
554 err:
555 kfree_skb(skb);
556 return err;
557 }
558
cxio_hal_destroy_ctrl_qp(struct cxio_rdev * rdev_p)559 static int cxio_hal_destroy_ctrl_qp(struct cxio_rdev *rdev_p)
560 {
561 dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
562 (1UL << T3_CTRL_QP_SIZE_LOG2)
563 * sizeof(union t3_wr), rdev_p->ctrl_qp.workq,
564 dma_unmap_addr(&rdev_p->ctrl_qp, mapping));
565 return cxio_hal_clear_qp_ctx(rdev_p, T3_CTRL_QP_ID);
566 }
567
568 /* write len bytes of data into addr (32B aligned address)
569 * If data is NULL, clear len byte of memory to zero.
570 * caller acquires the ctrl_qp lock before the call
571 */
cxio_hal_ctrl_qp_write_mem(struct cxio_rdev * rdev_p,u32 addr,u32 len,void * data)572 static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev *rdev_p, u32 addr,
573 u32 len, void *data)
574 {
575 u32 i, nr_wqe, copy_len;
576 u8 *copy_data;
577 u8 wr_len, utx_len; /* length in 8 byte flit */
578 enum t3_wr_flags flag;
579 __be64 *wqe;
580 u64 utx_cmd;
581 addr &= 0x7FFFFFF;
582 nr_wqe = len % 96 ? len / 96 + 1 : len / 96; /* 96B max per WQE */
583 pr_debug("%s wptr 0x%x rptr 0x%x len %d, nr_wqe %d data %p addr 0x%0x\n",
584 __func__, rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, len,
585 nr_wqe, data, addr);
586 utx_len = 3; /* in 32B unit */
587 for (i = 0; i < nr_wqe; i++) {
588 if (Q_FULL(rdev_p->ctrl_qp.rptr, rdev_p->ctrl_qp.wptr,
589 T3_CTRL_QP_SIZE_LOG2)) {
590 pr_debug("%s ctrl_qp full wtpr 0x%0x rptr 0x%0x, wait for more space i %d\n",
591 __func__,
592 rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, i);
593 if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
594 !Q_FULL(rdev_p->ctrl_qp.rptr,
595 rdev_p->ctrl_qp.wptr,
596 T3_CTRL_QP_SIZE_LOG2))) {
597 pr_debug("%s ctrl_qp workq interrupted\n",
598 __func__);
599 return -ERESTARTSYS;
600 }
601 pr_debug("%s ctrl_qp wakeup, continue posting work request i %d\n",
602 __func__, i);
603 }
604 wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
605 (1 << T3_CTRL_QP_SIZE_LOG2)));
606 flag = 0;
607 if (i == (nr_wqe - 1)) {
608 /* last WQE */
609 flag = T3_COMPLETION_FLAG;
610 if (len % 32)
611 utx_len = len / 32 + 1;
612 else
613 utx_len = len / 32;
614 }
615
616 /*
617 * Force a CQE to return the credit to the workq in case
618 * we posted more than half the max QP size of WRs
619 */
620 if ((i != 0) &&
621 (i % (((1 << T3_CTRL_QP_SIZE_LOG2)) >> 1) == 0)) {
622 flag = T3_COMPLETION_FLAG;
623 pr_debug("%s force completion at i %d\n", __func__, i);
624 }
625
626 /* build the utx mem command */
627 wqe += (sizeof(struct t3_bypass_wr) >> 3);
628 utx_cmd = (T3_UTX_MEM_WRITE << 28) | (addr + i * 3);
629 utx_cmd <<= 32;
630 utx_cmd |= (utx_len << 28) | ((utx_len << 2) + 1);
631 *wqe = cpu_to_be64(utx_cmd);
632 wqe++;
633 copy_data = (u8 *) data + i * 96;
634 copy_len = len > 96 ? 96 : len;
635
636 /* clear memory content if data is NULL */
637 if (data)
638 memcpy(wqe, copy_data, copy_len);
639 else
640 memset(wqe, 0, copy_len);
641 if (copy_len % 32)
642 memset(((u8 *) wqe) + copy_len, 0,
643 32 - (copy_len % 32));
644 wr_len = ((sizeof(struct t3_bypass_wr)) >> 3) + 1 +
645 (utx_len << 2);
646 wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
647 (1 << T3_CTRL_QP_SIZE_LOG2)));
648
649 /* wptr in the WRID[31:0] */
650 ((union t3_wrid *)(wqe+1))->id0.low = rdev_p->ctrl_qp.wptr;
651
652 /*
653 * This must be the last write with a memory barrier
654 * for the genbit
655 */
656 build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_BP, flag,
657 Q_GENBIT(rdev_p->ctrl_qp.wptr,
658 T3_CTRL_QP_SIZE_LOG2), T3_CTRL_QP_ID,
659 wr_len, T3_SOPEOP);
660 if (flag == T3_COMPLETION_FLAG)
661 ring_doorbell(rdev_p->ctrl_qp.doorbell, T3_CTRL_QP_ID);
662 len -= 96;
663 rdev_p->ctrl_qp.wptr++;
664 }
665 return 0;
666 }
667
668 /* IN: stag key, pdid, perm, zbva, to, len, page_size, pbl_size and pbl_addr
669 * OUT: stag index
670 * TBD: shared memory region support
671 */
__cxio_tpt_op(struct cxio_rdev * rdev_p,u32 reset_tpt_entry,u32 * stag,u8 stag_state,u32 pdid,enum tpt_mem_type type,enum tpt_mem_perm perm,u32 zbva,u64 to,u32 len,u8 page_size,u32 pbl_size,u32 pbl_addr)672 static int __cxio_tpt_op(struct cxio_rdev *rdev_p, u32 reset_tpt_entry,
673 u32 *stag, u8 stag_state, u32 pdid,
674 enum tpt_mem_type type, enum tpt_mem_perm perm,
675 u32 zbva, u64 to, u32 len, u8 page_size,
676 u32 pbl_size, u32 pbl_addr)
677 {
678 int err;
679 struct tpt_entry tpt;
680 u32 stag_idx;
681 u32 wptr;
682
683 if (cxio_fatal_error(rdev_p))
684 return -EIO;
685
686 stag_state = stag_state > 0;
687 stag_idx = (*stag) >> 8;
688
689 if ((!reset_tpt_entry) && !(*stag != T3_STAG_UNSET)) {
690 stag_idx = cxio_hal_get_stag(rdev_p->rscp);
691 if (!stag_idx)
692 return -ENOMEM;
693 *stag = (stag_idx << 8) | ((*stag) & 0xFF);
694 }
695 pr_debug("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
696 __func__, stag_state, type, pdid, stag_idx);
697
698 mutex_lock(&rdev_p->ctrl_qp.lock);
699
700 /* write TPT entry */
701 if (reset_tpt_entry)
702 memset(&tpt, 0, sizeof(tpt));
703 else {
704 tpt.valid_stag_pdid = cpu_to_be32(F_TPT_VALID |
705 V_TPT_STAG_KEY((*stag) & M_TPT_STAG_KEY) |
706 V_TPT_STAG_STATE(stag_state) |
707 V_TPT_STAG_TYPE(type) | V_TPT_PDID(pdid));
708 BUG_ON(page_size >= 28);
709 tpt.flags_pagesize_qpid = cpu_to_be32(V_TPT_PERM(perm) |
710 ((perm & TPT_MW_BIND) ? F_TPT_MW_BIND_ENABLE : 0) |
711 V_TPT_ADDR_TYPE((zbva ? TPT_ZBTO : TPT_VATO)) |
712 V_TPT_PAGE_SIZE(page_size));
713 tpt.rsvd_pbl_addr = cpu_to_be32(V_TPT_PBL_ADDR(PBL_OFF(rdev_p, pbl_addr)>>3));
714 tpt.len = cpu_to_be32(len);
715 tpt.va_hi = cpu_to_be32((u32) (to >> 32));
716 tpt.va_low_or_fbo = cpu_to_be32((u32) (to & 0xFFFFFFFFULL));
717 tpt.rsvd_bind_cnt_or_pstag = 0;
718 tpt.rsvd_pbl_size = cpu_to_be32(V_TPT_PBL_SIZE(pbl_size >> 2));
719 }
720 err = cxio_hal_ctrl_qp_write_mem(rdev_p,
721 stag_idx +
722 (rdev_p->rnic_info.tpt_base >> 5),
723 sizeof(tpt), &tpt);
724
725 /* release the stag index to free pool */
726 if (reset_tpt_entry)
727 cxio_hal_put_stag(rdev_p->rscp, stag_idx);
728
729 wptr = rdev_p->ctrl_qp.wptr;
730 mutex_unlock(&rdev_p->ctrl_qp.lock);
731 if (!err)
732 if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
733 SEQ32_GE(rdev_p->ctrl_qp.rptr,
734 wptr)))
735 return -ERESTARTSYS;
736 return err;
737 }
738
cxio_write_pbl(struct cxio_rdev * rdev_p,__be64 * pbl,u32 pbl_addr,u32 pbl_size)739 int cxio_write_pbl(struct cxio_rdev *rdev_p, __be64 *pbl,
740 u32 pbl_addr, u32 pbl_size)
741 {
742 u32 wptr;
743 int err;
744
745 pr_debug("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
746 __func__, pbl_addr, rdev_p->rnic_info.pbl_base,
747 pbl_size);
748
749 mutex_lock(&rdev_p->ctrl_qp.lock);
750 err = cxio_hal_ctrl_qp_write_mem(rdev_p, pbl_addr >> 5, pbl_size << 3,
751 pbl);
752 wptr = rdev_p->ctrl_qp.wptr;
753 mutex_unlock(&rdev_p->ctrl_qp.lock);
754 if (err)
755 return err;
756
757 if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
758 SEQ32_GE(rdev_p->ctrl_qp.rptr,
759 wptr)))
760 return -ERESTARTSYS;
761
762 return 0;
763 }
764
cxio_register_phys_mem(struct cxio_rdev * rdev_p,u32 * stag,u32 pdid,enum tpt_mem_perm perm,u32 zbva,u64 to,u32 len,u8 page_size,u32 pbl_size,u32 pbl_addr)765 int cxio_register_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
766 enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
767 u8 page_size, u32 pbl_size, u32 pbl_addr)
768 {
769 *stag = T3_STAG_UNSET;
770 return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
771 zbva, to, len, page_size, pbl_size, pbl_addr);
772 }
773
cxio_reregister_phys_mem(struct cxio_rdev * rdev_p,u32 * stag,u32 pdid,enum tpt_mem_perm perm,u32 zbva,u64 to,u32 len,u8 page_size,u32 pbl_size,u32 pbl_addr)774 int cxio_reregister_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
775 enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
776 u8 page_size, u32 pbl_size, u32 pbl_addr)
777 {
778 return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
779 zbva, to, len, page_size, pbl_size, pbl_addr);
780 }
781
cxio_dereg_mem(struct cxio_rdev * rdev_p,u32 stag,u32 pbl_size,u32 pbl_addr)782 int cxio_dereg_mem(struct cxio_rdev *rdev_p, u32 stag, u32 pbl_size,
783 u32 pbl_addr)
784 {
785 return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
786 pbl_size, pbl_addr);
787 }
788
cxio_allocate_window(struct cxio_rdev * rdev_p,u32 * stag,u32 pdid)789 int cxio_allocate_window(struct cxio_rdev *rdev_p, u32 * stag, u32 pdid)
790 {
791 *stag = T3_STAG_UNSET;
792 return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_MW, 0, 0, 0ULL, 0, 0,
793 0, 0);
794 }
795
cxio_deallocate_window(struct cxio_rdev * rdev_p,u32 stag)796 int cxio_deallocate_window(struct cxio_rdev *rdev_p, u32 stag)
797 {
798 return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
799 0, 0);
800 }
801
cxio_allocate_stag(struct cxio_rdev * rdev_p,u32 * stag,u32 pdid,u32 pbl_size,u32 pbl_addr)802 int cxio_allocate_stag(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid, u32 pbl_size, u32 pbl_addr)
803 {
804 *stag = T3_STAG_UNSET;
805 return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_NON_SHARED_MR,
806 0, 0, 0ULL, 0, 0, pbl_size, pbl_addr);
807 }
808
cxio_rdma_init(struct cxio_rdev * rdev_p,struct t3_rdma_init_attr * attr)809 int cxio_rdma_init(struct cxio_rdev *rdev_p, struct t3_rdma_init_attr *attr)
810 {
811 struct t3_rdma_init_wr *wqe;
812 struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_ATOMIC);
813 if (!skb)
814 return -ENOMEM;
815 pr_debug("%s rdev_p %p\n", __func__, rdev_p);
816 wqe = __skb_put(skb, sizeof(*wqe));
817 wqe->wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_INIT));
818 wqe->wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(attr->tid) |
819 V_FW_RIWR_LEN(sizeof(*wqe) >> 3));
820 wqe->wrid.id1 = 0;
821 wqe->qpid = cpu_to_be32(attr->qpid);
822 wqe->pdid = cpu_to_be32(attr->pdid);
823 wqe->scqid = cpu_to_be32(attr->scqid);
824 wqe->rcqid = cpu_to_be32(attr->rcqid);
825 wqe->rq_addr = cpu_to_be32(attr->rq_addr - rdev_p->rnic_info.rqt_base);
826 wqe->rq_size = cpu_to_be32(attr->rq_size);
827 wqe->mpaattrs = attr->mpaattrs;
828 wqe->qpcaps = attr->qpcaps;
829 wqe->ulpdu_size = cpu_to_be16(attr->tcp_emss);
830 wqe->rqe_count = cpu_to_be16(attr->rqe_count);
831 wqe->flags_rtr_type = cpu_to_be16(attr->flags |
832 V_RTR_TYPE(attr->rtr_type) |
833 V_CHAN(attr->chan));
834 wqe->ord = cpu_to_be32(attr->ord);
835 wqe->ird = cpu_to_be32(attr->ird);
836 wqe->qp_dma_addr = cpu_to_be64(attr->qp_dma_addr);
837 wqe->qp_dma_size = cpu_to_be32(attr->qp_dma_size);
838 wqe->irs = cpu_to_be32(attr->irs);
839 skb->priority = 0; /* 0=>ToeQ; 1=>CtrlQ */
840 return iwch_cxgb3_ofld_send(rdev_p->t3cdev_p, skb);
841 }
842
cxio_register_ev_cb(cxio_hal_ev_callback_func_t ev_cb)843 void cxio_register_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
844 {
845 cxio_ev_cb = ev_cb;
846 }
847
cxio_unregister_ev_cb(cxio_hal_ev_callback_func_t ev_cb)848 void cxio_unregister_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
849 {
850 cxio_ev_cb = NULL;
851 }
852
cxio_hal_ev_handler(struct t3cdev * t3cdev_p,struct sk_buff * skb)853 static int cxio_hal_ev_handler(struct t3cdev *t3cdev_p, struct sk_buff *skb)
854 {
855 static int cnt;
856 struct cxio_rdev *rdev_p = NULL;
857 struct respQ_msg_t *rsp_msg = (struct respQ_msg_t *) skb->data;
858 pr_debug("%d: %s cq_id 0x%x cq_ptr 0x%x genbit %0x overflow %0x an %0x se %0x notify %0x cqbranch %0x creditth %0x\n",
859 cnt, __func__, RSPQ_CQID(rsp_msg), RSPQ_CQPTR(rsp_msg),
860 RSPQ_GENBIT(rsp_msg), RSPQ_OVERFLOW(rsp_msg), RSPQ_AN(rsp_msg),
861 RSPQ_SE(rsp_msg), RSPQ_NOTIFY(rsp_msg), RSPQ_CQBRANCH(rsp_msg),
862 RSPQ_CREDIT_THRESH(rsp_msg));
863 pr_debug("CQE: QPID 0x%0x genbit %0x type 0x%0x status 0x%0x opcode %d len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
864 CQE_QPID(rsp_msg->cqe), CQE_GENBIT(rsp_msg->cqe),
865 CQE_TYPE(rsp_msg->cqe), CQE_STATUS(rsp_msg->cqe),
866 CQE_OPCODE(rsp_msg->cqe), CQE_LEN(rsp_msg->cqe),
867 CQE_WRID_HI(rsp_msg->cqe), CQE_WRID_LOW(rsp_msg->cqe));
868 rdev_p = (struct cxio_rdev *)t3cdev_p->ulp;
869 if (!rdev_p) {
870 pr_debug("%s called by t3cdev %p with null ulp\n", __func__,
871 t3cdev_p);
872 return 0;
873 }
874 if (CQE_QPID(rsp_msg->cqe) == T3_CTRL_QP_ID) {
875 rdev_p->ctrl_qp.rptr = CQE_WRID_LOW(rsp_msg->cqe) + 1;
876 wake_up_interruptible(&rdev_p->ctrl_qp.waitq);
877 dev_kfree_skb_irq(skb);
878 } else if (CQE_QPID(rsp_msg->cqe) == 0xfff8)
879 dev_kfree_skb_irq(skb);
880 else if (cxio_ev_cb)
881 (*cxio_ev_cb) (rdev_p, skb);
882 else
883 dev_kfree_skb_irq(skb);
884 cnt++;
885 return 0;
886 }
887
888 /* Caller takes care of locking if needed */
cxio_rdev_open(struct cxio_rdev * rdev_p)889 int cxio_rdev_open(struct cxio_rdev *rdev_p)
890 {
891 struct net_device *netdev_p = NULL;
892 int err = 0;
893 if (strlen(rdev_p->dev_name)) {
894 if (cxio_hal_find_rdev_by_name(rdev_p->dev_name)) {
895 return -EBUSY;
896 }
897 netdev_p = dev_get_by_name(&init_net, rdev_p->dev_name);
898 if (!netdev_p) {
899 return -EINVAL;
900 }
901 dev_put(netdev_p);
902 } else if (rdev_p->t3cdev_p) {
903 if (cxio_hal_find_rdev_by_t3cdev(rdev_p->t3cdev_p)) {
904 return -EBUSY;
905 }
906 netdev_p = rdev_p->t3cdev_p->lldev;
907 strncpy(rdev_p->dev_name, rdev_p->t3cdev_p->name,
908 T3_MAX_DEV_NAME_LEN);
909 } else {
910 pr_debug("%s t3cdev_p or dev_name must be set\n", __func__);
911 return -EINVAL;
912 }
913
914 list_add_tail(&rdev_p->entry, &rdev_list);
915
916 pr_debug("%s opening rnic dev %s\n", __func__, rdev_p->dev_name);
917 memset(&rdev_p->ctrl_qp, 0, sizeof(rdev_p->ctrl_qp));
918 if (!rdev_p->t3cdev_p)
919 rdev_p->t3cdev_p = dev2t3cdev(netdev_p);
920 rdev_p->t3cdev_p->ulp = (void *) rdev_p;
921
922 err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, GET_EMBEDDED_INFO,
923 &(rdev_p->fw_info));
924 if (err) {
925 pr_err("%s t3cdev_p(%p)->ctl returned error %d\n",
926 __func__, rdev_p->t3cdev_p, err);
927 goto err1;
928 }
929 if (G_FW_VERSION_MAJOR(rdev_p->fw_info.fw_vers) != CXIO_FW_MAJ) {
930 pr_err("fatal firmware version mismatch: need version %u but adapter has version %u\n",
931 CXIO_FW_MAJ,
932 G_FW_VERSION_MAJOR(rdev_p->fw_info.fw_vers));
933 err = -EINVAL;
934 goto err1;
935 }
936
937 err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_GET_PARAMS,
938 &(rdev_p->rnic_info));
939 if (err) {
940 pr_err("%s t3cdev_p(%p)->ctl returned error %d\n",
941 __func__, rdev_p->t3cdev_p, err);
942 goto err1;
943 }
944 err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, GET_PORTS,
945 &(rdev_p->port_info));
946 if (err) {
947 pr_err("%s t3cdev_p(%p)->ctl returned error %d\n",
948 __func__, rdev_p->t3cdev_p, err);
949 goto err1;
950 }
951
952 /*
953 * qpshift is the number of bits to shift the qpid left in order
954 * to get the correct address of the doorbell for that qp.
955 */
956 cxio_init_ucontext(rdev_p, &rdev_p->uctx);
957 rdev_p->qpshift = PAGE_SHIFT -
958 ilog2(65536 >>
959 ilog2(rdev_p->rnic_info.udbell_len >>
960 PAGE_SHIFT));
961 rdev_p->qpnr = rdev_p->rnic_info.udbell_len >> PAGE_SHIFT;
962 rdev_p->qpmask = (65536 >> ilog2(rdev_p->qpnr)) - 1;
963 pr_debug("%s rnic %s info: tpt_base 0x%0x tpt_top 0x%0x num stags %d pbl_base 0x%0x pbl_top 0x%0x rqt_base 0x%0x, rqt_top 0x%0x\n",
964 __func__, rdev_p->dev_name, rdev_p->rnic_info.tpt_base,
965 rdev_p->rnic_info.tpt_top, cxio_num_stags(rdev_p),
966 rdev_p->rnic_info.pbl_base,
967 rdev_p->rnic_info.pbl_top, rdev_p->rnic_info.rqt_base,
968 rdev_p->rnic_info.rqt_top);
969 pr_debug("udbell_len 0x%0x udbell_physbase 0x%lx kdb_addr %p qpshift %lu qpnr %d qpmask 0x%x\n",
970 rdev_p->rnic_info.udbell_len,
971 rdev_p->rnic_info.udbell_physbase, rdev_p->rnic_info.kdb_addr,
972 rdev_p->qpshift, rdev_p->qpnr, rdev_p->qpmask);
973
974 err = cxio_hal_init_ctrl_qp(rdev_p);
975 if (err) {
976 pr_err("%s error %d initializing ctrl_qp\n", __func__, err);
977 goto err1;
978 }
979 err = cxio_hal_init_resource(rdev_p, cxio_num_stags(rdev_p), 0,
980 0, T3_MAX_NUM_QP, T3_MAX_NUM_CQ,
981 T3_MAX_NUM_PD);
982 if (err) {
983 pr_err("%s error %d initializing hal resources\n",
984 __func__, err);
985 goto err2;
986 }
987 err = cxio_hal_pblpool_create(rdev_p);
988 if (err) {
989 pr_err("%s error %d initializing pbl mem pool\n",
990 __func__, err);
991 goto err3;
992 }
993 err = cxio_hal_rqtpool_create(rdev_p);
994 if (err) {
995 pr_err("%s error %d initializing rqt mem pool\n",
996 __func__, err);
997 goto err4;
998 }
999 return 0;
1000 err4:
1001 cxio_hal_pblpool_destroy(rdev_p);
1002 err3:
1003 cxio_hal_destroy_resource(rdev_p->rscp);
1004 err2:
1005 cxio_hal_destroy_ctrl_qp(rdev_p);
1006 err1:
1007 rdev_p->t3cdev_p->ulp = NULL;
1008 list_del(&rdev_p->entry);
1009 return err;
1010 }
1011
cxio_rdev_close(struct cxio_rdev * rdev_p)1012 void cxio_rdev_close(struct cxio_rdev *rdev_p)
1013 {
1014 if (rdev_p) {
1015 cxio_hal_pblpool_destroy(rdev_p);
1016 cxio_hal_rqtpool_destroy(rdev_p);
1017 list_del(&rdev_p->entry);
1018 cxio_hal_destroy_ctrl_qp(rdev_p);
1019 cxio_hal_destroy_resource(rdev_p->rscp);
1020 rdev_p->t3cdev_p->ulp = NULL;
1021 }
1022 }
1023
cxio_hal_init(void)1024 int __init cxio_hal_init(void)
1025 {
1026 if (cxio_hal_init_rhdl_resource(T3_MAX_NUM_RI))
1027 return -ENOMEM;
1028 t3_register_cpl_handler(CPL_ASYNC_NOTIF, cxio_hal_ev_handler);
1029 return 0;
1030 }
1031
cxio_hal_exit(void)1032 void __exit cxio_hal_exit(void)
1033 {
1034 struct cxio_rdev *rdev, *tmp;
1035
1036 t3_register_cpl_handler(CPL_ASYNC_NOTIF, NULL);
1037 list_for_each_entry_safe(rdev, tmp, &rdev_list, entry)
1038 cxio_rdev_close(rdev);
1039 cxio_hal_destroy_rhdl_resource();
1040 }
1041
flush_completed_wrs(struct t3_wq * wq,struct t3_cq * cq)1042 static void flush_completed_wrs(struct t3_wq *wq, struct t3_cq *cq)
1043 {
1044 struct t3_swsq *sqp;
1045 __u32 ptr = wq->sq_rptr;
1046 int count = Q_COUNT(wq->sq_rptr, wq->sq_wptr);
1047
1048 sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
1049 while (count--)
1050 if (!sqp->signaled) {
1051 ptr++;
1052 sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
1053 } else if (sqp->complete) {
1054
1055 /*
1056 * Insert this completed cqe into the swcq.
1057 */
1058 pr_debug("%s moving cqe into swcq sq idx %ld cq idx %ld\n",
1059 __func__, Q_PTR2IDX(ptr, wq->sq_size_log2),
1060 Q_PTR2IDX(cq->sw_wptr, cq->size_log2));
1061 sqp->cqe.header |= htonl(V_CQE_SWCQE(1));
1062 *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2))
1063 = sqp->cqe;
1064 cq->sw_wptr++;
1065 sqp->signaled = 0;
1066 break;
1067 } else
1068 break;
1069 }
1070
create_read_req_cqe(struct t3_wq * wq,struct t3_cqe * hw_cqe,struct t3_cqe * read_cqe)1071 static void create_read_req_cqe(struct t3_wq *wq, struct t3_cqe *hw_cqe,
1072 struct t3_cqe *read_cqe)
1073 {
1074 read_cqe->u.scqe.wrid_hi = wq->oldest_read->sq_wptr;
1075 read_cqe->len = wq->oldest_read->read_len;
1076 read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(*hw_cqe)) |
1077 V_CQE_SWCQE(SW_CQE(*hw_cqe)) |
1078 V_CQE_OPCODE(T3_READ_REQ) |
1079 V_CQE_TYPE(1));
1080 }
1081
1082 /*
1083 * Return a ptr to the next read wr in the SWSQ or NULL.
1084 */
advance_oldest_read(struct t3_wq * wq)1085 static void advance_oldest_read(struct t3_wq *wq)
1086 {
1087
1088 u32 rptr = wq->oldest_read - wq->sq + 1;
1089 u32 wptr = Q_PTR2IDX(wq->sq_wptr, wq->sq_size_log2);
1090
1091 while (Q_PTR2IDX(rptr, wq->sq_size_log2) != wptr) {
1092 wq->oldest_read = wq->sq + Q_PTR2IDX(rptr, wq->sq_size_log2);
1093
1094 if (wq->oldest_read->opcode == T3_READ_REQ)
1095 return;
1096 rptr++;
1097 }
1098 wq->oldest_read = NULL;
1099 }
1100
1101 /*
1102 * cxio_poll_cq
1103 *
1104 * Caller must:
1105 * check the validity of the first CQE,
1106 * supply the wq assicated with the qpid.
1107 *
1108 * credit: cq credit to return to sge.
1109 * cqe_flushed: 1 iff the CQE is flushed.
1110 * cqe: copy of the polled CQE.
1111 *
1112 * return value:
1113 * 0 CQE returned,
1114 * -1 CQE skipped, try again.
1115 */
cxio_poll_cq(struct t3_wq * wq,struct t3_cq * cq,struct t3_cqe * cqe,u8 * cqe_flushed,u64 * cookie,u32 * credit)1116 int cxio_poll_cq(struct t3_wq *wq, struct t3_cq *cq, struct t3_cqe *cqe,
1117 u8 *cqe_flushed, u64 *cookie, u32 *credit)
1118 {
1119 int ret = 0;
1120 struct t3_cqe *hw_cqe, read_cqe;
1121
1122 *cqe_flushed = 0;
1123 *credit = 0;
1124 hw_cqe = cxio_next_cqe(cq);
1125
1126 pr_debug("%s CQE OOO %d qpid 0x%0x genbit %d type %d status 0x%0x opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
1127 __func__, CQE_OOO(*hw_cqe), CQE_QPID(*hw_cqe),
1128 CQE_GENBIT(*hw_cqe), CQE_TYPE(*hw_cqe), CQE_STATUS(*hw_cqe),
1129 CQE_OPCODE(*hw_cqe), CQE_LEN(*hw_cqe), CQE_WRID_HI(*hw_cqe),
1130 CQE_WRID_LOW(*hw_cqe));
1131
1132 /*
1133 * skip cqe's not affiliated with a QP.
1134 */
1135 if (wq == NULL) {
1136 ret = -1;
1137 goto skip_cqe;
1138 }
1139
1140 /*
1141 * Gotta tweak READ completions:
1142 * 1) the cqe doesn't contain the sq_wptr from the wr.
1143 * 2) opcode not reflected from the wr.
1144 * 3) read_len not reflected from the wr.
1145 * 4) cq_type is RQ_TYPE not SQ_TYPE.
1146 */
1147 if (RQ_TYPE(*hw_cqe) && (CQE_OPCODE(*hw_cqe) == T3_READ_RESP)) {
1148
1149 /*
1150 * If this is an unsolicited read response, then the read
1151 * was generated by the kernel driver as part of peer-2-peer
1152 * connection setup. So ignore the completion.
1153 */
1154 if (!wq->oldest_read) {
1155 if (CQE_STATUS(*hw_cqe))
1156 wq->error = 1;
1157 ret = -1;
1158 goto skip_cqe;
1159 }
1160
1161 /*
1162 * Don't write to the HWCQ, so create a new read req CQE
1163 * in local memory.
1164 */
1165 create_read_req_cqe(wq, hw_cqe, &read_cqe);
1166 hw_cqe = &read_cqe;
1167 advance_oldest_read(wq);
1168 }
1169
1170 /*
1171 * T3A: Discard TERMINATE CQEs.
1172 */
1173 if (CQE_OPCODE(*hw_cqe) == T3_TERMINATE) {
1174 ret = -1;
1175 wq->error = 1;
1176 goto skip_cqe;
1177 }
1178
1179 if (CQE_STATUS(*hw_cqe) || wq->error) {
1180 *cqe_flushed = wq->error;
1181 wq->error = 1;
1182
1183 /*
1184 * T3A inserts errors into the CQE. We cannot return
1185 * these as work completions.
1186 */
1187 /* incoming write failures */
1188 if ((CQE_OPCODE(*hw_cqe) == T3_RDMA_WRITE)
1189 && RQ_TYPE(*hw_cqe)) {
1190 ret = -1;
1191 goto skip_cqe;
1192 }
1193 /* incoming read request failures */
1194 if ((CQE_OPCODE(*hw_cqe) == T3_READ_RESP) && SQ_TYPE(*hw_cqe)) {
1195 ret = -1;
1196 goto skip_cqe;
1197 }
1198
1199 /* incoming SEND with no receive posted failures */
1200 if (CQE_SEND_OPCODE(*hw_cqe) && RQ_TYPE(*hw_cqe) &&
1201 Q_EMPTY(wq->rq_rptr, wq->rq_wptr)) {
1202 ret = -1;
1203 goto skip_cqe;
1204 }
1205 BUG_ON((*cqe_flushed == 0) && !SW_CQE(*hw_cqe));
1206 goto proc_cqe;
1207 }
1208
1209 /*
1210 * RECV completion.
1211 */
1212 if (RQ_TYPE(*hw_cqe)) {
1213
1214 /*
1215 * HW only validates 4 bits of MSN. So we must validate that
1216 * the MSN in the SEND is the next expected MSN. If its not,
1217 * then we complete this with TPT_ERR_MSN and mark the wq in
1218 * error.
1219 */
1220
1221 if (Q_EMPTY(wq->rq_rptr, wq->rq_wptr)) {
1222 wq->error = 1;
1223 ret = -1;
1224 goto skip_cqe;
1225 }
1226
1227 if (unlikely((CQE_WRID_MSN(*hw_cqe) != (wq->rq_rptr + 1)))) {
1228 wq->error = 1;
1229 hw_cqe->header |= htonl(V_CQE_STATUS(TPT_ERR_MSN));
1230 goto proc_cqe;
1231 }
1232 goto proc_cqe;
1233 }
1234
1235 /*
1236 * If we get here its a send completion.
1237 *
1238 * Handle out of order completion. These get stuffed
1239 * in the SW SQ. Then the SW SQ is walked to move any
1240 * now in-order completions into the SW CQ. This handles
1241 * 2 cases:
1242 * 1) reaping unsignaled WRs when the first subsequent
1243 * signaled WR is completed.
1244 * 2) out of order read completions.
1245 */
1246 if (!SW_CQE(*hw_cqe) && (CQE_WRID_SQ_WPTR(*hw_cqe) != wq->sq_rptr)) {
1247 struct t3_swsq *sqp;
1248
1249 pr_debug("%s out of order completion going in swsq at idx %ld\n",
1250 __func__,
1251 Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe),
1252 wq->sq_size_log2));
1253 sqp = wq->sq +
1254 Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2);
1255 sqp->cqe = *hw_cqe;
1256 sqp->complete = 1;
1257 ret = -1;
1258 goto flush_wq;
1259 }
1260
1261 proc_cqe:
1262 *cqe = *hw_cqe;
1263
1264 /*
1265 * Reap the associated WR(s) that are freed up with this
1266 * completion.
1267 */
1268 if (SQ_TYPE(*hw_cqe)) {
1269 wq->sq_rptr = CQE_WRID_SQ_WPTR(*hw_cqe);
1270 pr_debug("%s completing sq idx %ld\n", __func__,
1271 Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2));
1272 *cookie = wq->sq[Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2)].wr_id;
1273 wq->sq_rptr++;
1274 } else {
1275 pr_debug("%s completing rq idx %ld\n", __func__,
1276 Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
1277 *cookie = wq->rq[Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2)].wr_id;
1278 if (wq->rq[Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2)].pbl_addr)
1279 cxio_hal_pblpool_free(wq->rdev,
1280 wq->rq[Q_PTR2IDX(wq->rq_rptr,
1281 wq->rq_size_log2)].pbl_addr, T3_STAG0_PBL_SIZE);
1282 BUG_ON(Q_EMPTY(wq->rq_rptr, wq->rq_wptr));
1283 wq->rq_rptr++;
1284 }
1285
1286 flush_wq:
1287 /*
1288 * Flush any completed cqes that are now in-order.
1289 */
1290 flush_completed_wrs(wq, cq);
1291
1292 skip_cqe:
1293 if (SW_CQE(*hw_cqe)) {
1294 pr_debug("%s cq %p cqid 0x%x skip sw cqe sw_rptr 0x%x\n",
1295 __func__, cq, cq->cqid, cq->sw_rptr);
1296 ++cq->sw_rptr;
1297 } else {
1298 pr_debug("%s cq %p cqid 0x%x skip hw cqe rptr 0x%x\n",
1299 __func__, cq, cq->cqid, cq->rptr);
1300 ++cq->rptr;
1301
1302 /*
1303 * T3A: compute credits.
1304 */
1305 if (((cq->rptr - cq->wptr) > (1 << (cq->size_log2 - 1)))
1306 || ((cq->rptr - cq->wptr) >= 128)) {
1307 *credit = cq->rptr - cq->wptr;
1308 cq->wptr = cq->rptr;
1309 }
1310 }
1311 return ret;
1312 }
1313