1
2
3
4
5
6#include <linux/vmalloc.h>
7
8#include <rdma/ib_addr.h>
9#include <rdma/ib_umem.h>
10#include <rdma/ib_user_verbs.h>
11#include <rdma/ib_verbs.h>
12#include <rdma/uverbs_ioctl.h>
13
14#include "efa.h"
15
16#define EFA_MMAP_FLAG_SHIFT 56
17#define EFA_MMAP_PAGE_MASK GENMASK(EFA_MMAP_FLAG_SHIFT - 1, 0)
18#define EFA_MMAP_INVALID U64_MAX
19
20enum {
21 EFA_MMAP_DMA_PAGE = 0,
22 EFA_MMAP_IO_WC,
23 EFA_MMAP_IO_NC,
24};
25
26#define EFA_AENQ_ENABLED_GROUPS \
27 (BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
28 BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
29
30struct efa_mmap_entry {
31 void *obj;
32 u64 address;
33 u64 length;
34 u32 mmap_page;
35 u8 mmap_flag;
36};
37
38static inline u64 get_mmap_key(const struct efa_mmap_entry *efa)
39{
40 return ((u64)efa->mmap_flag << EFA_MMAP_FLAG_SHIFT) |
41 ((u64)efa->mmap_page << PAGE_SHIFT);
42}
43
44#define EFA_CHUNK_PAYLOAD_SHIFT 12
45#define EFA_CHUNK_PAYLOAD_SIZE BIT(EFA_CHUNK_PAYLOAD_SHIFT)
46#define EFA_CHUNK_PAYLOAD_PTR_SIZE 8
47
48#define EFA_CHUNK_SHIFT 12
49#define EFA_CHUNK_SIZE BIT(EFA_CHUNK_SHIFT)
50#define EFA_CHUNK_PTR_SIZE sizeof(struct efa_com_ctrl_buff_info)
51
52#define EFA_PTRS_PER_CHUNK \
53 ((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
54
55#define EFA_CHUNK_USED_SIZE \
56 ((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
57
58#define EFA_SUPPORTED_ACCESS_FLAGS IB_ACCESS_LOCAL_WRITE
59
60struct pbl_chunk {
61 dma_addr_t dma_addr;
62 u64 *buf;
63 u32 length;
64};
65
66struct pbl_chunk_list {
67 struct pbl_chunk *chunks;
68 unsigned int size;
69};
70
71struct pbl_context {
72 union {
73 struct {
74 dma_addr_t dma_addr;
75 } continuous;
76 struct {
77 u32 pbl_buf_size_in_pages;
78 struct scatterlist *sgl;
79 int sg_dma_cnt;
80 struct pbl_chunk_list chunk_list;
81 } indirect;
82 } phys;
83 u64 *pbl_buf;
84 u32 pbl_buf_size_in_bytes;
85 u8 physically_continuous;
86};
87
88static inline struct efa_dev *to_edev(struct ib_device *ibdev)
89{
90 return container_of(ibdev, struct efa_dev, ibdev);
91}
92
93static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
94{
95 return container_of(ibucontext, struct efa_ucontext, ibucontext);
96}
97
98static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
99{
100 return container_of(ibpd, struct efa_pd, ibpd);
101}
102
103static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
104{
105 return container_of(ibmr, struct efa_mr, ibmr);
106}
107
108static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
109{
110 return container_of(ibqp, struct efa_qp, ibqp);
111}
112
113static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
114{
115 return container_of(ibcq, struct efa_cq, ibcq);
116}
117
118static inline struct efa_ah *to_eah(struct ib_ah *ibah)
119{
120 return container_of(ibah, struct efa_ah, ibah);
121}
122
123#define field_avail(x, fld, sz) (offsetof(typeof(x), fld) + \
124 sizeof(((typeof(x) *)0)->fld) <= (sz))
125
126#define is_reserved_cleared(reserved) \
127 !memchr_inv(reserved, 0, sizeof(reserved))
128
129static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
130 size_t size, enum dma_data_direction dir)
131{
132 void *addr;
133
134 addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
135 if (!addr)
136 return NULL;
137
138 *dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
139 if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
140 ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
141 free_pages_exact(addr, size);
142 return NULL;
143 }
144
145 return addr;
146}
147
148
149
150
151
152
153
154
155static void mmap_entries_remove_free(struct efa_dev *dev,
156 struct efa_ucontext *ucontext)
157{
158 struct efa_mmap_entry *entry;
159 unsigned long mmap_page;
160
161 xa_for_each(&ucontext->mmap_xa, mmap_page, entry) {
162 xa_erase(&ucontext->mmap_xa, mmap_page);
163
164 ibdev_dbg(
165 &dev->ibdev,
166 "mmap: obj[0x%p] key[%#llx] addr[%#llx] len[%#llx] removed\n",
167 entry->obj, get_mmap_key(entry), entry->address,
168 entry->length);
169 if (entry->mmap_flag == EFA_MMAP_DMA_PAGE)
170
171 free_pages_exact(phys_to_virt(entry->address),
172 entry->length);
173 kfree(entry);
174 }
175}
176
177static struct efa_mmap_entry *mmap_entry_get(struct efa_dev *dev,
178 struct efa_ucontext *ucontext,
179 u64 key, u64 len)
180{
181 struct efa_mmap_entry *entry;
182 u64 mmap_page;
183
184 mmap_page = (key & EFA_MMAP_PAGE_MASK) >> PAGE_SHIFT;
185 if (mmap_page > U32_MAX)
186 return NULL;
187
188 entry = xa_load(&ucontext->mmap_xa, mmap_page);
189 if (!entry || get_mmap_key(entry) != key || entry->length != len)
190 return NULL;
191
192 ibdev_dbg(&dev->ibdev,
193 "mmap: obj[0x%p] key[%#llx] addr[%#llx] len[%#llx] removed\n",
194 entry->obj, key, entry->address, entry->length);
195
196 return entry;
197}
198
199
200
201
202
203static u64 mmap_entry_insert(struct efa_dev *dev, struct efa_ucontext *ucontext,
204 void *obj, u64 address, u64 length, u8 mmap_flag)
205{
206 struct efa_mmap_entry *entry;
207 u32 next_mmap_page;
208 int err;
209
210 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
211 if (!entry)
212 return EFA_MMAP_INVALID;
213
214 entry->obj = obj;
215 entry->address = address;
216 entry->length = length;
217 entry->mmap_flag = mmap_flag;
218
219 xa_lock(&ucontext->mmap_xa);
220 if (check_add_overflow(ucontext->mmap_xa_page,
221 (u32)(length >> PAGE_SHIFT),
222 &next_mmap_page))
223 goto err_unlock;
224
225 entry->mmap_page = ucontext->mmap_xa_page;
226 ucontext->mmap_xa_page = next_mmap_page;
227 err = __xa_insert(&ucontext->mmap_xa, entry->mmap_page, entry,
228 GFP_KERNEL);
229 if (err)
230 goto err_unlock;
231
232 xa_unlock(&ucontext->mmap_xa);
233
234 ibdev_dbg(
235 &dev->ibdev,
236 "mmap: obj[0x%p] addr[%#llx], len[%#llx], key[%#llx] inserted\n",
237 entry->obj, entry->address, entry->length, get_mmap_key(entry));
238
239 return get_mmap_key(entry);
240
241err_unlock:
242 xa_unlock(&ucontext->mmap_xa);
243 kfree(entry);
244 return EFA_MMAP_INVALID;
245
246}
247
248int efa_query_device(struct ib_device *ibdev,
249 struct ib_device_attr *props,
250 struct ib_udata *udata)
251{
252 struct efa_com_get_device_attr_result *dev_attr;
253 struct efa_ibv_ex_query_device_resp resp = {};
254 struct efa_dev *dev = to_edev(ibdev);
255 int err;
256
257 if (udata && udata->inlen &&
258 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
259 ibdev_dbg(ibdev,
260 "Incompatible ABI params, udata not cleared\n");
261 return -EINVAL;
262 }
263
264 dev_attr = &dev->dev_attr;
265
266 memset(props, 0, sizeof(*props));
267 props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
268 props->page_size_cap = dev_attr->page_size_cap;
269 props->vendor_id = dev->pdev->vendor;
270 props->vendor_part_id = dev->pdev->device;
271 props->hw_ver = dev->pdev->subsystem_device;
272 props->max_qp = dev_attr->max_qp;
273 props->max_cq = dev_attr->max_cq;
274 props->max_pd = dev_attr->max_pd;
275 props->max_mr = dev_attr->max_mr;
276 props->max_ah = dev_attr->max_ah;
277 props->max_cqe = dev_attr->max_cq_depth;
278 props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
279 dev_attr->max_rq_depth);
280 props->max_send_sge = dev_attr->max_sq_sge;
281 props->max_recv_sge = dev_attr->max_rq_sge;
282
283 if (udata && udata->outlen) {
284 resp.max_sq_sge = dev_attr->max_sq_sge;
285 resp.max_rq_sge = dev_attr->max_rq_sge;
286 resp.max_sq_wr = dev_attr->max_sq_depth;
287 resp.max_rq_wr = dev_attr->max_rq_depth;
288
289 err = ib_copy_to_udata(udata, &resp,
290 min(sizeof(resp), udata->outlen));
291 if (err) {
292 ibdev_dbg(ibdev,
293 "Failed to copy udata for query_device\n");
294 return err;
295 }
296 }
297
298 return 0;
299}
300
301int efa_query_port(struct ib_device *ibdev, u8 port,
302 struct ib_port_attr *props)
303{
304 struct efa_dev *dev = to_edev(ibdev);
305
306 props->lmc = 1;
307
308 props->state = IB_PORT_ACTIVE;
309 props->phys_state = 5;
310 props->gid_tbl_len = 1;
311 props->pkey_tbl_len = 1;
312 props->active_speed = IB_SPEED_EDR;
313 props->active_width = IB_WIDTH_4X;
314 props->max_mtu = ib_mtu_int_to_enum(dev->mtu);
315 props->active_mtu = ib_mtu_int_to_enum(dev->mtu);
316 props->max_msg_sz = dev->mtu;
317 props->max_vl_num = 1;
318
319 return 0;
320}
321
322int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
323 int qp_attr_mask,
324 struct ib_qp_init_attr *qp_init_attr)
325{
326 struct efa_dev *dev = to_edev(ibqp->device);
327 struct efa_com_query_qp_params params = {};
328 struct efa_com_query_qp_result result;
329 struct efa_qp *qp = to_eqp(ibqp);
330 int err;
331
332#define EFA_QUERY_QP_SUPP_MASK \
333 (IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
334 IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP)
335
336 if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
337 ibdev_dbg(&dev->ibdev,
338 "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
339 qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
340 return -EOPNOTSUPP;
341 }
342
343 memset(qp_attr, 0, sizeof(*qp_attr));
344 memset(qp_init_attr, 0, sizeof(*qp_init_attr));
345
346 params.qp_handle = qp->qp_handle;
347 err = efa_com_query_qp(&dev->edev, ¶ms, &result);
348 if (err)
349 return err;
350
351 qp_attr->qp_state = result.qp_state;
352 qp_attr->qkey = result.qkey;
353 qp_attr->sq_psn = result.sq_psn;
354 qp_attr->sq_draining = result.sq_draining;
355 qp_attr->port_num = 1;
356
357 qp_attr->cap.max_send_wr = qp->max_send_wr;
358 qp_attr->cap.max_recv_wr = qp->max_recv_wr;
359 qp_attr->cap.max_send_sge = qp->max_send_sge;
360 qp_attr->cap.max_recv_sge = qp->max_recv_sge;
361 qp_attr->cap.max_inline_data = qp->max_inline_data;
362
363 qp_init_attr->qp_type = ibqp->qp_type;
364 qp_init_attr->recv_cq = ibqp->recv_cq;
365 qp_init_attr->send_cq = ibqp->send_cq;
366 qp_init_attr->qp_context = ibqp->qp_context;
367 qp_init_attr->cap = qp_attr->cap;
368
369 return 0;
370}
371
372int efa_query_gid(struct ib_device *ibdev, u8 port, int index,
373 union ib_gid *gid)
374{
375 struct efa_dev *dev = to_edev(ibdev);
376
377 memcpy(gid->raw, dev->addr, sizeof(dev->addr));
378
379 return 0;
380}
381
382int efa_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
383 u16 *pkey)
384{
385 if (index > 0)
386 return -EINVAL;
387
388 *pkey = 0xffff;
389 return 0;
390}
391
392static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
393{
394 struct efa_com_dealloc_pd_params params = {
395 .pdn = pdn,
396 };
397
398 return efa_com_dealloc_pd(&dev->edev, ¶ms);
399}
400
401int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
402{
403 struct efa_dev *dev = to_edev(ibpd->device);
404 struct efa_ibv_alloc_pd_resp resp = {};
405 struct efa_com_alloc_pd_result result;
406 struct efa_pd *pd = to_epd(ibpd);
407 int err;
408
409 if (udata->inlen &&
410 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
411 ibdev_dbg(&dev->ibdev,
412 "Incompatible ABI params, udata not cleared\n");
413 err = -EINVAL;
414 goto err_out;
415 }
416
417 err = efa_com_alloc_pd(&dev->edev, &result);
418 if (err)
419 goto err_out;
420
421 pd->pdn = result.pdn;
422 resp.pdn = result.pdn;
423
424 if (udata->outlen) {
425 err = ib_copy_to_udata(udata, &resp,
426 min(sizeof(resp), udata->outlen));
427 if (err) {
428 ibdev_dbg(&dev->ibdev,
429 "Failed to copy udata for alloc_pd\n");
430 goto err_dealloc_pd;
431 }
432 }
433
434 ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
435
436 return 0;
437
438err_dealloc_pd:
439 efa_pd_dealloc(dev, result.pdn);
440err_out:
441 atomic64_inc(&dev->stats.sw_stats.alloc_pd_err);
442 return err;
443}
444
445void efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
446{
447 struct efa_dev *dev = to_edev(ibpd->device);
448 struct efa_pd *pd = to_epd(ibpd);
449
450 if (udata->inlen &&
451 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
452 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
453 return;
454 }
455
456 ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
457 efa_pd_dealloc(dev, pd->pdn);
458}
459
460static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
461{
462 struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
463
464 return efa_com_destroy_qp(&dev->edev, ¶ms);
465}
466
467int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
468{
469 struct efa_dev *dev = to_edev(ibqp->pd->device);
470 struct efa_qp *qp = to_eqp(ibqp);
471 int err;
472
473 if (udata->inlen &&
474 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
475 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
476 return -EINVAL;
477 }
478
479 ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
480 err = efa_destroy_qp_handle(dev, qp->qp_handle);
481 if (err)
482 return err;
483
484 if (qp->rq_cpu_addr) {
485 ibdev_dbg(&dev->ibdev,
486 "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
487 qp->rq_cpu_addr, qp->rq_size,
488 &qp->rq_dma_addr);
489 dma_unmap_single(&dev->pdev->dev, qp->rq_dma_addr, qp->rq_size,
490 DMA_TO_DEVICE);
491 }
492
493 kfree(qp);
494 return 0;
495}
496
497static int qp_mmap_entries_setup(struct efa_qp *qp,
498 struct efa_dev *dev,
499 struct efa_ucontext *ucontext,
500 struct efa_com_create_qp_params *params,
501 struct efa_ibv_create_qp_resp *resp)
502{
503
504
505
506
507 resp->sq_db_mmap_key =
508 mmap_entry_insert(dev, ucontext, qp,
509 dev->db_bar_addr + resp->sq_db_offset,
510 PAGE_SIZE, EFA_MMAP_IO_NC);
511 if (resp->sq_db_mmap_key == EFA_MMAP_INVALID)
512 return -ENOMEM;
513
514 resp->sq_db_offset &= ~PAGE_MASK;
515
516 resp->llq_desc_mmap_key =
517 mmap_entry_insert(dev, ucontext, qp,
518 dev->mem_bar_addr + resp->llq_desc_offset,
519 PAGE_ALIGN(params->sq_ring_size_in_bytes +
520 (resp->llq_desc_offset & ~PAGE_MASK)),
521 EFA_MMAP_IO_WC);
522 if (resp->llq_desc_mmap_key == EFA_MMAP_INVALID)
523 return -ENOMEM;
524
525 resp->llq_desc_offset &= ~PAGE_MASK;
526
527 if (qp->rq_size) {
528 resp->rq_db_mmap_key =
529 mmap_entry_insert(dev, ucontext, qp,
530 dev->db_bar_addr + resp->rq_db_offset,
531 PAGE_SIZE, EFA_MMAP_IO_NC);
532 if (resp->rq_db_mmap_key == EFA_MMAP_INVALID)
533 return -ENOMEM;
534
535 resp->rq_db_offset &= ~PAGE_MASK;
536
537 resp->rq_mmap_key =
538 mmap_entry_insert(dev, ucontext, qp,
539 virt_to_phys(qp->rq_cpu_addr),
540 qp->rq_size, EFA_MMAP_DMA_PAGE);
541 if (resp->rq_mmap_key == EFA_MMAP_INVALID)
542 return -ENOMEM;
543
544 resp->rq_mmap_size = qp->rq_size;
545 }
546
547 return 0;
548}
549
550static int efa_qp_validate_cap(struct efa_dev *dev,
551 struct ib_qp_init_attr *init_attr)
552{
553 if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
554 ibdev_dbg(&dev->ibdev,
555 "qp: requested send wr[%u] exceeds the max[%u]\n",
556 init_attr->cap.max_send_wr,
557 dev->dev_attr.max_sq_depth);
558 return -EINVAL;
559 }
560 if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
561 ibdev_dbg(&dev->ibdev,
562 "qp: requested receive wr[%u] exceeds the max[%u]\n",
563 init_attr->cap.max_recv_wr,
564 dev->dev_attr.max_rq_depth);
565 return -EINVAL;
566 }
567 if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
568 ibdev_dbg(&dev->ibdev,
569 "qp: requested sge send[%u] exceeds the max[%u]\n",
570 init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
571 return -EINVAL;
572 }
573 if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
574 ibdev_dbg(&dev->ibdev,
575 "qp: requested sge recv[%u] exceeds the max[%u]\n",
576 init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
577 return -EINVAL;
578 }
579 if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
580 ibdev_dbg(&dev->ibdev,
581 "qp: requested inline data[%u] exceeds the max[%u]\n",
582 init_attr->cap.max_inline_data,
583 dev->dev_attr.inline_buf_size);
584 return -EINVAL;
585 }
586
587 return 0;
588}
589
590static int efa_qp_validate_attr(struct efa_dev *dev,
591 struct ib_qp_init_attr *init_attr)
592{
593 if (init_attr->qp_type != IB_QPT_DRIVER &&
594 init_attr->qp_type != IB_QPT_UD) {
595 ibdev_dbg(&dev->ibdev,
596 "Unsupported qp type %d\n", init_attr->qp_type);
597 return -EOPNOTSUPP;
598 }
599
600 if (init_attr->srq) {
601 ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
602 return -EOPNOTSUPP;
603 }
604
605 if (init_attr->create_flags) {
606 ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
607 return -EOPNOTSUPP;
608 }
609
610 return 0;
611}
612
613struct ib_qp *efa_create_qp(struct ib_pd *ibpd,
614 struct ib_qp_init_attr *init_attr,
615 struct ib_udata *udata)
616{
617 struct efa_com_create_qp_params create_qp_params = {};
618 struct efa_com_create_qp_result create_qp_resp;
619 struct efa_dev *dev = to_edev(ibpd->device);
620 struct efa_ibv_create_qp_resp resp = {};
621 struct efa_ibv_create_qp cmd = {};
622 bool rq_entry_inserted = false;
623 struct efa_ucontext *ucontext;
624 struct efa_qp *qp;
625 int err;
626
627 ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
628 ibucontext);
629
630 err = efa_qp_validate_cap(dev, init_attr);
631 if (err)
632 goto err_out;
633
634 err = efa_qp_validate_attr(dev, init_attr);
635 if (err)
636 goto err_out;
637
638 if (!field_avail(cmd, driver_qp_type, udata->inlen)) {
639 ibdev_dbg(&dev->ibdev,
640 "Incompatible ABI params, no input udata\n");
641 err = -EINVAL;
642 goto err_out;
643 }
644
645 if (udata->inlen > sizeof(cmd) &&
646 !ib_is_udata_cleared(udata, sizeof(cmd),
647 udata->inlen - sizeof(cmd))) {
648 ibdev_dbg(&dev->ibdev,
649 "Incompatible ABI params, unknown fields in udata\n");
650 err = -EINVAL;
651 goto err_out;
652 }
653
654 err = ib_copy_from_udata(&cmd, udata,
655 min(sizeof(cmd), udata->inlen));
656 if (err) {
657 ibdev_dbg(&dev->ibdev,
658 "Cannot copy udata for create_qp\n");
659 goto err_out;
660 }
661
662 if (cmd.comp_mask) {
663 ibdev_dbg(&dev->ibdev,
664 "Incompatible ABI params, unknown fields in udata\n");
665 err = -EINVAL;
666 goto err_out;
667 }
668
669 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
670 if (!qp) {
671 err = -ENOMEM;
672 goto err_out;
673 }
674
675 create_qp_params.uarn = ucontext->uarn;
676 create_qp_params.pd = to_epd(ibpd)->pdn;
677
678 if (init_attr->qp_type == IB_QPT_UD) {
679 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
680 } else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
681 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
682 } else {
683 ibdev_dbg(&dev->ibdev,
684 "Unsupported qp type %d driver qp type %d\n",
685 init_attr->qp_type, cmd.driver_qp_type);
686 err = -EOPNOTSUPP;
687 goto err_free_qp;
688 }
689
690 ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
691 init_attr->qp_type, cmd.driver_qp_type);
692 create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
693 create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
694 create_qp_params.sq_depth = init_attr->cap.max_send_wr;
695 create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
696
697 create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
698 create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
699 qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
700 if (qp->rq_size) {
701 qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
702 qp->rq_size, DMA_TO_DEVICE);
703 if (!qp->rq_cpu_addr) {
704 err = -ENOMEM;
705 goto err_free_qp;
706 }
707
708 ibdev_dbg(&dev->ibdev,
709 "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
710 qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
711 create_qp_params.rq_base_addr = qp->rq_dma_addr;
712 }
713
714 err = efa_com_create_qp(&dev->edev, &create_qp_params,
715 &create_qp_resp);
716 if (err)
717 goto err_free_mapped;
718
719 resp.sq_db_offset = create_qp_resp.sq_db_offset;
720 resp.rq_db_offset = create_qp_resp.rq_db_offset;
721 resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
722 resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
723 resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
724
725 err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
726 &resp);
727 if (err)
728 goto err_destroy_qp;
729
730 rq_entry_inserted = true;
731 qp->qp_handle = create_qp_resp.qp_handle;
732 qp->ibqp.qp_num = create_qp_resp.qp_num;
733 qp->ibqp.qp_type = init_attr->qp_type;
734 qp->max_send_wr = init_attr->cap.max_send_wr;
735 qp->max_recv_wr = init_attr->cap.max_recv_wr;
736 qp->max_send_sge = init_attr->cap.max_send_sge;
737 qp->max_recv_sge = init_attr->cap.max_recv_sge;
738 qp->max_inline_data = init_attr->cap.max_inline_data;
739
740 if (udata->outlen) {
741 err = ib_copy_to_udata(udata, &resp,
742 min(sizeof(resp), udata->outlen));
743 if (err) {
744 ibdev_dbg(&dev->ibdev,
745 "Failed to copy udata for qp[%u]\n",
746 create_qp_resp.qp_num);
747 goto err_destroy_qp;
748 }
749 }
750
751 ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
752
753 return &qp->ibqp;
754
755err_destroy_qp:
756 efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
757err_free_mapped:
758 if (qp->rq_size) {
759 dma_unmap_single(&dev->pdev->dev, qp->rq_dma_addr, qp->rq_size,
760 DMA_TO_DEVICE);
761 if (!rq_entry_inserted)
762 free_pages_exact(qp->rq_cpu_addr, qp->rq_size);
763 }
764err_free_qp:
765 kfree(qp);
766err_out:
767 atomic64_inc(&dev->stats.sw_stats.create_qp_err);
768 return ERR_PTR(err);
769}
770
771static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
772 struct ib_qp_attr *qp_attr, int qp_attr_mask,
773 enum ib_qp_state cur_state,
774 enum ib_qp_state new_state)
775{
776#define EFA_MODIFY_QP_SUPP_MASK \
777 (IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
778 IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN)
779
780 if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
781 ibdev_dbg(&dev->ibdev,
782 "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
783 qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
784 return -EOPNOTSUPP;
785 }
786
787 if (!ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
788 qp_attr_mask)) {
789 ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
790 return -EINVAL;
791 }
792
793 if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
794 ibdev_dbg(&dev->ibdev, "Can't change port num\n");
795 return -EOPNOTSUPP;
796 }
797
798 if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
799 ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
800 return -EOPNOTSUPP;
801 }
802
803 return 0;
804}
805
806int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
807 int qp_attr_mask, struct ib_udata *udata)
808{
809 struct efa_dev *dev = to_edev(ibqp->device);
810 struct efa_com_modify_qp_params params = {};
811 struct efa_qp *qp = to_eqp(ibqp);
812 enum ib_qp_state cur_state;
813 enum ib_qp_state new_state;
814 int err;
815
816 if (udata->inlen &&
817 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
818 ibdev_dbg(&dev->ibdev,
819 "Incompatible ABI params, udata not cleared\n");
820 return -EINVAL;
821 }
822
823 cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
824 qp->state;
825 new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
826
827 err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
828 new_state);
829 if (err)
830 return err;
831
832 params.qp_handle = qp->qp_handle;
833
834 if (qp_attr_mask & IB_QP_STATE) {
835 params.modify_mask |= BIT(EFA_ADMIN_QP_STATE_BIT) |
836 BIT(EFA_ADMIN_CUR_QP_STATE_BIT);
837 params.cur_qp_state = qp_attr->cur_qp_state;
838 params.qp_state = qp_attr->qp_state;
839 }
840
841 if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
842 params.modify_mask |=
843 BIT(EFA_ADMIN_SQ_DRAINED_ASYNC_NOTIFY_BIT);
844 params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
845 }
846
847 if (qp_attr_mask & IB_QP_QKEY) {
848 params.modify_mask |= BIT(EFA_ADMIN_QKEY_BIT);
849 params.qkey = qp_attr->qkey;
850 }
851
852 if (qp_attr_mask & IB_QP_SQ_PSN) {
853 params.modify_mask |= BIT(EFA_ADMIN_SQ_PSN_BIT);
854 params.sq_psn = qp_attr->sq_psn;
855 }
856
857 err = efa_com_modify_qp(&dev->edev, ¶ms);
858 if (err)
859 return err;
860
861 qp->state = new_state;
862
863 return 0;
864}
865
866static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
867{
868 struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
869
870 return efa_com_destroy_cq(&dev->edev, ¶ms);
871}
872
873int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
874{
875 struct efa_dev *dev = to_edev(ibcq->device);
876 struct efa_cq *cq = to_ecq(ibcq);
877 int err;
878
879 if (udata->inlen &&
880 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
881 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
882 return -EINVAL;
883 }
884
885 ibdev_dbg(&dev->ibdev,
886 "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
887 cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
888
889 err = efa_destroy_cq_idx(dev, cq->cq_idx);
890 if (err)
891 return err;
892
893 dma_unmap_single(&dev->pdev->dev, cq->dma_addr, cq->size,
894 DMA_FROM_DEVICE);
895
896 kfree(cq);
897 return 0;
898}
899
900static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
901 struct efa_ibv_create_cq_resp *resp)
902{
903 resp->q_mmap_size = cq->size;
904 resp->q_mmap_key = mmap_entry_insert(dev, cq->ucontext, cq,
905 virt_to_phys(cq->cpu_addr),
906 cq->size, EFA_MMAP_DMA_PAGE);
907 if (resp->q_mmap_key == EFA_MMAP_INVALID)
908 return -ENOMEM;
909
910 return 0;
911}
912
913static struct ib_cq *do_create_cq(struct ib_device *ibdev, int entries,
914 int vector, struct ib_ucontext *ibucontext,
915 struct ib_udata *udata)
916{
917 struct efa_ibv_create_cq_resp resp = {};
918 struct efa_com_create_cq_params params;
919 struct efa_com_create_cq_result result;
920 struct efa_dev *dev = to_edev(ibdev);
921 struct efa_ibv_create_cq cmd = {};
922 bool cq_entry_inserted = false;
923 struct efa_cq *cq;
924 int err;
925
926 ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
927
928 if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
929 ibdev_dbg(ibdev,
930 "cq: requested entries[%u] non-positive or greater than max[%u]\n",
931 entries, dev->dev_attr.max_cq_depth);
932 err = -EINVAL;
933 goto err_out;
934 }
935
936 if (!field_avail(cmd, num_sub_cqs, udata->inlen)) {
937 ibdev_dbg(ibdev,
938 "Incompatible ABI params, no input udata\n");
939 err = -EINVAL;
940 goto err_out;
941 }
942
943 if (udata->inlen > sizeof(cmd) &&
944 !ib_is_udata_cleared(udata, sizeof(cmd),
945 udata->inlen - sizeof(cmd))) {
946 ibdev_dbg(ibdev,
947 "Incompatible ABI params, unknown fields in udata\n");
948 err = -EINVAL;
949 goto err_out;
950 }
951
952 err = ib_copy_from_udata(&cmd, udata,
953 min(sizeof(cmd), udata->inlen));
954 if (err) {
955 ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
956 goto err_out;
957 }
958
959 if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_50)) {
960 ibdev_dbg(ibdev,
961 "Incompatible ABI params, unknown fields in udata\n");
962 err = -EINVAL;
963 goto err_out;
964 }
965
966 if (!cmd.cq_entry_size) {
967 ibdev_dbg(ibdev,
968 "Invalid entry size [%u]\n", cmd.cq_entry_size);
969 err = -EINVAL;
970 goto err_out;
971 }
972
973 if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
974 ibdev_dbg(ibdev,
975 "Invalid number of sub cqs[%u] expected[%u]\n",
976 cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
977 err = -EINVAL;
978 goto err_out;
979 }
980
981 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
982 if (!cq) {
983 err = -ENOMEM;
984 goto err_out;
985 }
986
987 cq->ucontext = to_eucontext(ibucontext);
988 cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
989 cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
990 DMA_FROM_DEVICE);
991 if (!cq->cpu_addr) {
992 err = -ENOMEM;
993 goto err_free_cq;
994 }
995
996 params.uarn = cq->ucontext->uarn;
997 params.cq_depth = entries;
998 params.dma_addr = cq->dma_addr;
999 params.entry_size_in_bytes = cmd.cq_entry_size;
1000 params.num_sub_cqs = cmd.num_sub_cqs;
1001 err = efa_com_create_cq(&dev->edev, ¶ms, &result);
1002 if (err)
1003 goto err_free_mapped;
1004
1005 resp.cq_idx = result.cq_idx;
1006 cq->cq_idx = result.cq_idx;
1007 cq->ibcq.cqe = result.actual_depth;
1008 WARN_ON_ONCE(entries != result.actual_depth);
1009
1010 err = cq_mmap_entries_setup(dev, cq, &resp);
1011 if (err) {
1012 ibdev_dbg(ibdev,
1013 "Could not setup cq[%u] mmap entries\n", cq->cq_idx);
1014 goto err_destroy_cq;
1015 }
1016
1017 cq_entry_inserted = true;
1018
1019 if (udata->outlen) {
1020 err = ib_copy_to_udata(udata, &resp,
1021 min(sizeof(resp), udata->outlen));
1022 if (err) {
1023 ibdev_dbg(ibdev,
1024 "Failed to copy udata for create_cq\n");
1025 goto err_destroy_cq;
1026 }
1027 }
1028
1029 ibdev_dbg(ibdev,
1030 "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
1031 cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
1032
1033 return &cq->ibcq;
1034
1035err_destroy_cq:
1036 efa_destroy_cq_idx(dev, cq->cq_idx);
1037err_free_mapped:
1038 dma_unmap_single(&dev->pdev->dev, cq->dma_addr, cq->size,
1039 DMA_FROM_DEVICE);
1040 if (!cq_entry_inserted)
1041 free_pages_exact(cq->cpu_addr, cq->size);
1042err_free_cq:
1043 kfree(cq);
1044err_out:
1045 atomic64_inc(&dev->stats.sw_stats.create_cq_err);
1046 return ERR_PTR(err);
1047}
1048
1049struct ib_cq *efa_create_cq(struct ib_device *ibdev,
1050 const struct ib_cq_init_attr *attr,
1051 struct ib_udata *udata)
1052{
1053 struct efa_ucontext *ucontext = rdma_udata_to_drv_context(udata,
1054 struct efa_ucontext,
1055 ibucontext);
1056
1057 return do_create_cq(ibdev, attr->cqe, attr->comp_vector,
1058 &ucontext->ibucontext, udata);
1059}
1060
1061static int umem_to_page_list(struct efa_dev *dev,
1062 struct ib_umem *umem,
1063 u64 *page_list,
1064 u32 hp_cnt,
1065 u8 hp_shift)
1066{
1067 u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT);
1068 struct sg_dma_page_iter sg_iter;
1069 unsigned int page_idx = 0;
1070 unsigned int hp_idx = 0;
1071
1072 ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n",
1073 hp_cnt, pages_in_hp);
1074
1075 for_each_sg_dma_page(umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
1076 if (page_idx % pages_in_hp == 0) {
1077 page_list[hp_idx] = sg_page_iter_dma_address(&sg_iter);
1078 hp_idx++;
1079 }
1080
1081 page_idx++;
1082 }
1083
1084 return 0;
1085}
1086
1087static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1088{
1089 struct scatterlist *sglist;
1090 struct page *pg;
1091 int i;
1092
1093 sglist = kcalloc(page_cnt, sizeof(*sglist), GFP_KERNEL);
1094 if (!sglist)
1095 return NULL;
1096 sg_init_table(sglist, page_cnt);
1097 for (i = 0; i < page_cnt; i++) {
1098 pg = vmalloc_to_page(buf);
1099 if (!pg)
1100 goto err;
1101 sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1102 buf += PAGE_SIZE / sizeof(*buf);
1103 }
1104 return sglist;
1105
1106err:
1107 kfree(sglist);
1108 return NULL;
1109}
1110
1111
1112
1113
1114
1115static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1116{
1117 unsigned int entry, payloads_in_sg, chunk_list_size, chunk_idx, payload_idx;
1118 struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1119 int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1120 struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1121 int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1122 struct efa_com_ctrl_buff_info *ctrl_buf;
1123 u64 *cur_chunk_buf, *prev_chunk_buf;
1124 struct scatterlist *sg;
1125 dma_addr_t dma_addr;
1126 int i;
1127
1128
1129 chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1130
1131 chunk_list->size = chunk_list_size;
1132 chunk_list->chunks = kcalloc(chunk_list_size,
1133 sizeof(*chunk_list->chunks),
1134 GFP_KERNEL);
1135 if (!chunk_list->chunks)
1136 return -ENOMEM;
1137
1138 ibdev_dbg(&dev->ibdev,
1139 "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1140 page_cnt);
1141
1142
1143 for (i = 0; i < chunk_list_size; i++) {
1144 chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1145 if (!chunk_list->chunks[i].buf)
1146 goto chunk_list_dealloc;
1147
1148 chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1149 }
1150 chunk_list->chunks[chunk_list_size - 1].length =
1151 ((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1152 EFA_CHUNK_PTR_SIZE;
1153
1154
1155 chunk_idx = 0;
1156 payload_idx = 0;
1157 cur_chunk_buf = chunk_list->chunks[0].buf;
1158 for_each_sg(pages_sgl, sg, sg_dma_cnt, entry) {
1159 payloads_in_sg = sg_dma_len(sg) >> EFA_CHUNK_PAYLOAD_SHIFT;
1160 for (i = 0; i < payloads_in_sg; i++) {
1161 cur_chunk_buf[payload_idx++] =
1162 (sg_dma_address(sg) & ~(EFA_CHUNK_PAYLOAD_SIZE - 1)) +
1163 (EFA_CHUNK_PAYLOAD_SIZE * i);
1164
1165 if (payload_idx == EFA_PTRS_PER_CHUNK) {
1166 chunk_idx++;
1167 cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1168 payload_idx = 0;
1169 }
1170 }
1171 }
1172
1173
1174 for (i = chunk_list_size - 1; i >= 0; i--) {
1175 dma_addr = dma_map_single(&dev->pdev->dev,
1176 chunk_list->chunks[i].buf,
1177 chunk_list->chunks[i].length,
1178 DMA_TO_DEVICE);
1179 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1180 ibdev_err(&dev->ibdev,
1181 "chunk[%u] dma_map_failed\n", i);
1182 goto chunk_list_unmap;
1183 }
1184
1185 chunk_list->chunks[i].dma_addr = dma_addr;
1186 ibdev_dbg(&dev->ibdev,
1187 "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1188
1189 if (!i)
1190 break;
1191
1192 prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1193
1194 ctrl_buf = (struct efa_com_ctrl_buff_info *)
1195 &prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1196 ctrl_buf->length = chunk_list->chunks[i].length;
1197
1198 efa_com_set_dma_addr(dma_addr,
1199 &ctrl_buf->address.mem_addr_high,
1200 &ctrl_buf->address.mem_addr_low);
1201 }
1202
1203 return 0;
1204
1205chunk_list_unmap:
1206 for (; i < chunk_list_size; i++) {
1207 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1208 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1209 }
1210chunk_list_dealloc:
1211 for (i = 0; i < chunk_list_size; i++)
1212 kfree(chunk_list->chunks[i].buf);
1213
1214 kfree(chunk_list->chunks);
1215 return -ENOMEM;
1216}
1217
1218static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1219{
1220 struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1221 int i;
1222
1223 for (i = 0; i < chunk_list->size; i++) {
1224 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1225 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1226 kfree(chunk_list->chunks[i].buf);
1227 }
1228
1229 kfree(chunk_list->chunks);
1230}
1231
1232
1233static int pbl_continuous_initialize(struct efa_dev *dev,
1234 struct pbl_context *pbl)
1235{
1236 dma_addr_t dma_addr;
1237
1238 dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1239 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1240 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1241 ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1242 return -ENOMEM;
1243 }
1244
1245 pbl->phys.continuous.dma_addr = dma_addr;
1246 ibdev_dbg(&dev->ibdev,
1247 "pbl continuous - dma_addr = %pad, size[%u]\n",
1248 &dma_addr, pbl->pbl_buf_size_in_bytes);
1249
1250 return 0;
1251}
1252
1253
1254
1255
1256
1257
1258static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1259{
1260 u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, PAGE_SIZE);
1261 struct scatterlist *sgl;
1262 int sg_dma_cnt, err;
1263
1264 BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1265 sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1266 if (!sgl)
1267 return -ENOMEM;
1268
1269 sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1270 if (!sg_dma_cnt) {
1271 err = -EINVAL;
1272 goto err_map;
1273 }
1274
1275 pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1276 pbl->phys.indirect.sgl = sgl;
1277 pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1278 err = pbl_chunk_list_create(dev, pbl);
1279 if (err) {
1280 ibdev_dbg(&dev->ibdev,
1281 "chunk_list creation failed[%d]\n", err);
1282 goto err_chunk;
1283 }
1284
1285 ibdev_dbg(&dev->ibdev,
1286 "pbl indirect - size[%u], chunks[%u]\n",
1287 pbl->pbl_buf_size_in_bytes,
1288 pbl->phys.indirect.chunk_list.size);
1289
1290 return 0;
1291
1292err_chunk:
1293 dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1294err_map:
1295 kfree(sgl);
1296 return err;
1297}
1298
1299static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1300{
1301 pbl_chunk_list_destroy(dev, pbl);
1302 dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1303 pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1304 kfree(pbl->phys.indirect.sgl);
1305}
1306
1307
1308static int pbl_create(struct efa_dev *dev,
1309 struct pbl_context *pbl,
1310 struct ib_umem *umem,
1311 int hp_cnt,
1312 u8 hp_shift)
1313{
1314 int err;
1315
1316 pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1317 pbl->pbl_buf = kzalloc(pbl->pbl_buf_size_in_bytes,
1318 GFP_KERNEL | __GFP_NOWARN);
1319 if (pbl->pbl_buf) {
1320 pbl->physically_continuous = 1;
1321 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1322 hp_shift);
1323 if (err)
1324 goto err_continuous;
1325 err = pbl_continuous_initialize(dev, pbl);
1326 if (err)
1327 goto err_continuous;
1328 } else {
1329 pbl->physically_continuous = 0;
1330 pbl->pbl_buf = vzalloc(pbl->pbl_buf_size_in_bytes);
1331 if (!pbl->pbl_buf)
1332 return -ENOMEM;
1333
1334 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1335 hp_shift);
1336 if (err)
1337 goto err_indirect;
1338 err = pbl_indirect_initialize(dev, pbl);
1339 if (err)
1340 goto err_indirect;
1341 }
1342
1343 ibdev_dbg(&dev->ibdev,
1344 "user_pbl_created: user_pages[%u], continuous[%u]\n",
1345 hp_cnt, pbl->physically_continuous);
1346
1347 return 0;
1348
1349err_continuous:
1350 kfree(pbl->pbl_buf);
1351 return err;
1352err_indirect:
1353 vfree(pbl->pbl_buf);
1354 return err;
1355}
1356
1357static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1358{
1359 if (pbl->physically_continuous) {
1360 dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1361 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1362 kfree(pbl->pbl_buf);
1363 } else {
1364 pbl_indirect_terminate(dev, pbl);
1365 vfree(pbl->pbl_buf);
1366 }
1367}
1368
1369static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1370 struct efa_com_reg_mr_params *params)
1371{
1372 int err;
1373
1374 params->inline_pbl = 1;
1375 err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1376 params->page_num, params->page_shift);
1377 if (err)
1378 return err;
1379
1380 ibdev_dbg(&dev->ibdev,
1381 "inline_pbl_array - pages[%u]\n", params->page_num);
1382
1383 return 0;
1384}
1385
1386static int efa_create_pbl(struct efa_dev *dev,
1387 struct pbl_context *pbl,
1388 struct efa_mr *mr,
1389 struct efa_com_reg_mr_params *params)
1390{
1391 int err;
1392
1393 err = pbl_create(dev, pbl, mr->umem, params->page_num,
1394 params->page_shift);
1395 if (err) {
1396 ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1397 return err;
1398 }
1399
1400 params->inline_pbl = 0;
1401 params->indirect = !pbl->physically_continuous;
1402 if (pbl->physically_continuous) {
1403 params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1404
1405 efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1406 ¶ms->pbl.pbl.address.mem_addr_high,
1407 ¶ms->pbl.pbl.address.mem_addr_low);
1408 } else {
1409 params->pbl.pbl.length =
1410 pbl->phys.indirect.chunk_list.chunks[0].length;
1411
1412 efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1413 ¶ms->pbl.pbl.address.mem_addr_high,
1414 ¶ms->pbl.pbl.address.mem_addr_low);
1415 }
1416
1417 return 0;
1418}
1419
1420static void efa_cont_pages(struct ib_umem *umem, u64 addr,
1421 unsigned long max_page_shift,
1422 int *count, u8 *shift, u32 *ncont)
1423{
1424 struct scatterlist *sg;
1425 u64 base = ~0, p = 0;
1426 unsigned long tmp;
1427 unsigned long m;
1428 u64 len, pfn;
1429 int i = 0;
1430 int entry;
1431
1432 addr = addr >> PAGE_SHIFT;
1433 tmp = (unsigned long)addr;
1434 m = find_first_bit(&tmp, BITS_PER_LONG);
1435 if (max_page_shift)
1436 m = min_t(unsigned long, max_page_shift - PAGE_SHIFT, m);
1437
1438 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
1439 len = DIV_ROUND_UP(sg_dma_len(sg), PAGE_SIZE);
1440 pfn = sg_dma_address(sg) >> PAGE_SHIFT;
1441 if (base + p != pfn) {
1442
1443
1444
1445
1446 tmp = (unsigned long)(pfn | p);
1447 if (!IS_ALIGNED(tmp, 1 << m))
1448 m = find_first_bit(&tmp, BITS_PER_LONG);
1449
1450 base = pfn;
1451 p = 0;
1452 }
1453
1454 p += len;
1455 i += len;
1456 }
1457
1458 if (i) {
1459 m = min_t(unsigned long, ilog2(roundup_pow_of_two(i)), m);
1460 *ncont = DIV_ROUND_UP(i, (1 << m));
1461 } else {
1462 m = 0;
1463 *ncont = 0;
1464 }
1465
1466 *shift = PAGE_SHIFT + m;
1467 *count = i;
1468}
1469
1470struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1471 u64 virt_addr, int access_flags,
1472 struct ib_udata *udata)
1473{
1474 struct efa_dev *dev = to_edev(ibpd->device);
1475 struct efa_com_reg_mr_params params = {};
1476 struct efa_com_reg_mr_result result = {};
1477 unsigned long max_page_shift;
1478 struct pbl_context pbl;
1479 struct efa_mr *mr;
1480 int inline_size;
1481 int npages;
1482 int err;
1483
1484 if (udata->inlen &&
1485 !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
1486 ibdev_dbg(&dev->ibdev,
1487 "Incompatible ABI params, udata not cleared\n");
1488 err = -EINVAL;
1489 goto err_out;
1490 }
1491
1492 if (access_flags & ~EFA_SUPPORTED_ACCESS_FLAGS) {
1493 ibdev_dbg(&dev->ibdev,
1494 "Unsupported access flags[%#x], supported[%#x]\n",
1495 access_flags, EFA_SUPPORTED_ACCESS_FLAGS);
1496 err = -EOPNOTSUPP;
1497 goto err_out;
1498 }
1499
1500 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1501 if (!mr) {
1502 err = -ENOMEM;
1503 goto err_out;
1504 }
1505
1506 mr->umem = ib_umem_get(udata, start, length, access_flags, 0);
1507 if (IS_ERR(mr->umem)) {
1508 err = PTR_ERR(mr->umem);
1509 ibdev_dbg(&dev->ibdev,
1510 "Failed to pin and map user space memory[%d]\n", err);
1511 goto err_free;
1512 }
1513
1514 params.pd = to_epd(ibpd)->pdn;
1515 params.iova = virt_addr;
1516 params.mr_length_in_bytes = length;
1517 params.permissions = access_flags & 0x1;
1518 max_page_shift = fls64(dev->dev_attr.page_size_cap);
1519
1520 efa_cont_pages(mr->umem, start, max_page_shift, &npages,
1521 ¶ms.page_shift, ¶ms.page_num);
1522 ibdev_dbg(&dev->ibdev,
1523 "start %#llx length %#llx npages %d params.page_shift %u params.page_num %u\n",
1524 start, length, npages, params.page_shift, params.page_num);
1525
1526 inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1527 if (params.page_num <= inline_size) {
1528 err = efa_create_inline_pbl(dev, mr, ¶ms);
1529 if (err)
1530 goto err_unmap;
1531
1532 err = efa_com_register_mr(&dev->edev, ¶ms, &result);
1533 if (err)
1534 goto err_unmap;
1535 } else {
1536 err = efa_create_pbl(dev, &pbl, mr, ¶ms);
1537 if (err)
1538 goto err_unmap;
1539
1540 err = efa_com_register_mr(&dev->edev, ¶ms, &result);
1541 pbl_destroy(dev, &pbl);
1542
1543 if (err)
1544 goto err_unmap;
1545 }
1546
1547 mr->ibmr.lkey = result.l_key;
1548 mr->ibmr.rkey = result.r_key;
1549 mr->ibmr.length = length;
1550 ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1551
1552 return &mr->ibmr;
1553
1554err_unmap:
1555 ib_umem_release(mr->umem);
1556err_free:
1557 kfree(mr);
1558err_out:
1559 atomic64_inc(&dev->stats.sw_stats.reg_mr_err);
1560 return ERR_PTR(err);
1561}
1562
1563int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1564{
1565 struct efa_dev *dev = to_edev(ibmr->device);
1566 struct efa_com_dereg_mr_params params;
1567 struct efa_mr *mr = to_emr(ibmr);
1568 int err;
1569
1570 if (udata->inlen &&
1571 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1572 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1573 return -EINVAL;
1574 }
1575
1576 ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1577
1578 if (mr->umem) {
1579 params.l_key = mr->ibmr.lkey;
1580 err = efa_com_dereg_mr(&dev->edev, ¶ms);
1581 if (err)
1582 return err;
1583 ib_umem_release(mr->umem);
1584 }
1585
1586 kfree(mr);
1587
1588 return 0;
1589}
1590
1591int efa_get_port_immutable(struct ib_device *ibdev, u8 port_num,
1592 struct ib_port_immutable *immutable)
1593{
1594 struct ib_port_attr attr;
1595 int err;
1596
1597 err = ib_query_port(ibdev, port_num, &attr);
1598 if (err) {
1599 ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1600 return err;
1601 }
1602
1603 immutable->pkey_tbl_len = attr.pkey_tbl_len;
1604 immutable->gid_tbl_len = attr.gid_tbl_len;
1605
1606 return 0;
1607}
1608
1609static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1610{
1611 struct efa_com_dealloc_uar_params params = {
1612 .uarn = uarn,
1613 };
1614
1615 return efa_com_dealloc_uar(&dev->edev, ¶ms);
1616}
1617
1618int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1619{
1620 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1621 struct efa_dev *dev = to_edev(ibucontext->device);
1622 struct efa_ibv_alloc_ucontext_resp resp = {};
1623 struct efa_com_alloc_uar_result result;
1624 int err;
1625
1626
1627
1628
1629
1630
1631 err = efa_com_alloc_uar(&dev->edev, &result);
1632 if (err)
1633 goto err_out;
1634
1635 ucontext->uarn = result.uarn;
1636 xa_init(&ucontext->mmap_xa);
1637
1638 resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1639 resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1640 resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1641 resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1642 resp.max_llq_size = dev->dev_attr.max_llq_size;
1643
1644 if (udata && udata->outlen) {
1645 err = ib_copy_to_udata(udata, &resp,
1646 min(sizeof(resp), udata->outlen));
1647 if (err)
1648 goto err_dealloc_uar;
1649 }
1650
1651 return 0;
1652
1653err_dealloc_uar:
1654 efa_dealloc_uar(dev, result.uarn);
1655err_out:
1656 atomic64_inc(&dev->stats.sw_stats.alloc_ucontext_err);
1657 return err;
1658}
1659
1660void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
1661{
1662 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1663 struct efa_dev *dev = to_edev(ibucontext->device);
1664
1665 mmap_entries_remove_free(dev, ucontext);
1666 efa_dealloc_uar(dev, ucontext->uarn);
1667}
1668
1669static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
1670 struct vm_area_struct *vma, u64 key, u64 length)
1671{
1672 struct efa_mmap_entry *entry;
1673 unsigned long va;
1674 u64 pfn;
1675 int err;
1676
1677 entry = mmap_entry_get(dev, ucontext, key, length);
1678 if (!entry) {
1679 ibdev_dbg(&dev->ibdev, "key[%#llx] does not have valid entry\n",
1680 key);
1681 return -EINVAL;
1682 }
1683
1684 ibdev_dbg(&dev->ibdev,
1685 "Mapping address[%#llx], length[%#llx], mmap_flag[%d]\n",
1686 entry->address, length, entry->mmap_flag);
1687
1688 pfn = entry->address >> PAGE_SHIFT;
1689 switch (entry->mmap_flag) {
1690 case EFA_MMAP_IO_NC:
1691 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, length,
1692 pgprot_noncached(vma->vm_page_prot));
1693 break;
1694 case EFA_MMAP_IO_WC:
1695 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, length,
1696 pgprot_writecombine(vma->vm_page_prot));
1697 break;
1698 case EFA_MMAP_DMA_PAGE:
1699 for (va = vma->vm_start; va < vma->vm_end;
1700 va += PAGE_SIZE, pfn++) {
1701 err = vm_insert_page(vma, va, pfn_to_page(pfn));
1702 if (err)
1703 break;
1704 }
1705 break;
1706 default:
1707 err = -EINVAL;
1708 }
1709
1710 if (err)
1711 ibdev_dbg(
1712 &dev->ibdev,
1713 "Couldn't mmap address[%#llx] length[%#llx] mmap_flag[%d] err[%d]\n",
1714 entry->address, length, entry->mmap_flag, err);
1715
1716 return err;
1717}
1718
1719int efa_mmap(struct ib_ucontext *ibucontext,
1720 struct vm_area_struct *vma)
1721{
1722 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1723 struct efa_dev *dev = to_edev(ibucontext->device);
1724 u64 length = vma->vm_end - vma->vm_start;
1725 u64 key = vma->vm_pgoff << PAGE_SHIFT;
1726
1727 ibdev_dbg(&dev->ibdev,
1728 "start %#lx, end %#lx, length = %#llx, key = %#llx\n",
1729 vma->vm_start, vma->vm_end, length, key);
1730
1731 if (length % PAGE_SIZE != 0 || !(vma->vm_flags & VM_SHARED)) {
1732 ibdev_dbg(&dev->ibdev,
1733 "length[%#llx] is not page size aligned[%#lx] or VM_SHARED is not set [%#lx]\n",
1734 length, PAGE_SIZE, vma->vm_flags);
1735 return -EINVAL;
1736 }
1737
1738 if (vma->vm_flags & VM_EXEC) {
1739 ibdev_dbg(&dev->ibdev, "Mapping executable pages is not permitted\n");
1740 return -EPERM;
1741 }
1742
1743 return __efa_mmap(dev, ucontext, vma, key, length);
1744}
1745
1746static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
1747{
1748 struct efa_com_destroy_ah_params params = {
1749 .ah = ah->ah,
1750 .pdn = to_epd(ah->ibah.pd)->pdn,
1751 };
1752
1753 return efa_com_destroy_ah(&dev->edev, ¶ms);
1754}
1755
1756int efa_create_ah(struct ib_ah *ibah,
1757 struct rdma_ah_attr *ah_attr,
1758 u32 flags,
1759 struct ib_udata *udata)
1760{
1761 struct efa_dev *dev = to_edev(ibah->device);
1762 struct efa_com_create_ah_params params = {};
1763 struct efa_ibv_create_ah_resp resp = {};
1764 struct efa_com_create_ah_result result;
1765 struct efa_ah *ah = to_eah(ibah);
1766 int err;
1767
1768 if (!(flags & RDMA_CREATE_AH_SLEEPABLE)) {
1769 ibdev_dbg(&dev->ibdev,
1770 "Create address handle is not supported in atomic context\n");
1771 err = -EOPNOTSUPP;
1772 goto err_out;
1773 }
1774
1775 if (udata->inlen &&
1776 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1777 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1778 err = -EINVAL;
1779 goto err_out;
1780 }
1781
1782 memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
1783 sizeof(params.dest_addr));
1784 params.pdn = to_epd(ibah->pd)->pdn;
1785 err = efa_com_create_ah(&dev->edev, ¶ms, &result);
1786 if (err)
1787 goto err_out;
1788
1789 memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
1790 ah->ah = result.ah;
1791
1792 resp.efa_address_handle = result.ah;
1793
1794 if (udata->outlen) {
1795 err = ib_copy_to_udata(udata, &resp,
1796 min(sizeof(resp), udata->outlen));
1797 if (err) {
1798 ibdev_dbg(&dev->ibdev,
1799 "Failed to copy udata for create_ah response\n");
1800 goto err_destroy_ah;
1801 }
1802 }
1803 ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
1804
1805 return 0;
1806
1807err_destroy_ah:
1808 efa_ah_destroy(dev, ah);
1809err_out:
1810 atomic64_inc(&dev->stats.sw_stats.create_ah_err);
1811 return err;
1812}
1813
1814void efa_destroy_ah(struct ib_ah *ibah, u32 flags)
1815{
1816 struct efa_dev *dev = to_edev(ibah->pd->device);
1817 struct efa_ah *ah = to_eah(ibah);
1818
1819 ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
1820
1821 if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
1822 ibdev_dbg(&dev->ibdev,
1823 "Destroy address handle is not supported in atomic context\n");
1824 return;
1825 }
1826
1827 efa_ah_destroy(dev, ah);
1828}
1829
1830enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
1831 u8 port_num)
1832{
1833 return IB_LINK_LAYER_UNSPECIFIED;
1834}
1835
1836
