?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
dm writecache: add cond_resched to loop in persistent_memory_claim()
[linux/fpc-iii.git] / drivers / infiniband / hw / efa / efa_verbs.c
blob3420c774248616cec1483c30e8037302ee501e32
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3 * Copyright 2018-2020 Amazon.com, Inc. or its affiliates. All rights reserved.
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 enum {
17 EFA_MMAP_DMA_PAGE = 0,
18 EFA_MMAP_IO_WC,
19 EFA_MMAP_IO_NC,
20 };
21
22 #define EFA_AENQ_ENABLED_GROUPS \
23 (BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
24 BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
25
26 struct efa_user_mmap_entry {
27 struct rdma_user_mmap_entry rdma_entry;
28 u64 address;
29 u8 mmap_flag;
30 };
31
32 #define EFA_DEFINE_STATS(op) \
33 op(EFA_TX_BYTES, "tx_bytes") \
34 op(EFA_TX_PKTS, "tx_pkts") \
35 op(EFA_RX_BYTES, "rx_bytes") \
36 op(EFA_RX_PKTS, "rx_pkts") \
37 op(EFA_RX_DROPS, "rx_drops") \
38 op(EFA_SUBMITTED_CMDS, "submitted_cmds") \
39 op(EFA_COMPLETED_CMDS, "completed_cmds") \
40 op(EFA_NO_COMPLETION_CMDS, "no_completion_cmds") \
41 op(EFA_KEEP_ALIVE_RCVD, "keep_alive_rcvd") \
42 op(EFA_ALLOC_PD_ERR, "alloc_pd_err") \
43 op(EFA_CREATE_QP_ERR, "create_qp_err") \
44 op(EFA_REG_MR_ERR, "reg_mr_err") \
45 op(EFA_ALLOC_UCONTEXT_ERR, "alloc_ucontext_err") \
46 op(EFA_CREATE_AH_ERR, "create_ah_err")
47
48 #define EFA_STATS_ENUM(ename, name) ename,
49 #define EFA_STATS_STR(ename, name) [ename] = name,
50
51 enum efa_hw_stats {
52 EFA_DEFINE_STATS(EFA_STATS_ENUM)
53 };
54
55 static const char *const efa_stats_names[] = {
56 EFA_DEFINE_STATS(EFA_STATS_STR)
57 };
58
59 #define EFA_CHUNK_PAYLOAD_SHIFT 12
60 #define EFA_CHUNK_PAYLOAD_SIZE BIT(EFA_CHUNK_PAYLOAD_SHIFT)
61 #define EFA_CHUNK_PAYLOAD_PTR_SIZE 8
62
63 #define EFA_CHUNK_SHIFT 12
64 #define EFA_CHUNK_SIZE BIT(EFA_CHUNK_SHIFT)
65 #define EFA_CHUNK_PTR_SIZE sizeof(struct efa_com_ctrl_buff_info)
66
67 #define EFA_PTRS_PER_CHUNK \
68 ((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
69
70 #define EFA_CHUNK_USED_SIZE \
71 ((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
72
73 struct pbl_chunk {
74 dma_addr_t dma_addr;
75 u64 *buf;
76 u32 length;
77 };
78
79 struct pbl_chunk_list {
80 struct pbl_chunk *chunks;
81 unsigned int size;
82 };
83
84 struct pbl_context {
85 union {
86 struct {
87 dma_addr_t dma_addr;
88 } continuous;
89 struct {
90 u32 pbl_buf_size_in_pages;
91 struct scatterlist *sgl;
92 int sg_dma_cnt;
93 struct pbl_chunk_list chunk_list;
94 } indirect;
95 } phys;
96 u64 *pbl_buf;
97 u32 pbl_buf_size_in_bytes;
98 u8 physically_continuous;
99 };
100
101 static inline struct efa_dev *to_edev(struct ib_device *ibdev)
102 {
103 return container_of(ibdev, struct efa_dev, ibdev);
104 }
105
106 static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
107 {
108 return container_of(ibucontext, struct efa_ucontext, ibucontext);
109 }
110
111 static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
112 {
113 return container_of(ibpd, struct efa_pd, ibpd);
114 }
115
116 static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
117 {
118 return container_of(ibmr, struct efa_mr, ibmr);
119 }
120
121 static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
122 {
123 return container_of(ibqp, struct efa_qp, ibqp);
124 }
125
126 static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
127 {
128 return container_of(ibcq, struct efa_cq, ibcq);
129 }
130
131 static inline struct efa_ah *to_eah(struct ib_ah *ibah)
132 {
133 return container_of(ibah, struct efa_ah, ibah);
134 }
135
136 static inline struct efa_user_mmap_entry *
137 to_emmap(struct rdma_user_mmap_entry *rdma_entry)
138 {
139 return container_of(rdma_entry, struct efa_user_mmap_entry, rdma_entry);
140 }
141
142 static inline bool is_rdma_read_cap(struct efa_dev *dev)
143 {
144 return dev->dev_attr.device_caps & EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_RDMA_READ_MASK;
145 }
146
147 #define is_reserved_cleared(reserved) \
148 !memchr_inv(reserved, 0, sizeof(reserved))
149
150 static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
151 size_t size, enum dma_data_direction dir)
152 {
153 void *addr;
154
155 addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
156 if (!addr)
157 return NULL;
158
159 *dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
160 if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
161 ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
162 free_pages_exact(addr, size);
163 return NULL;
164 }
165
166 return addr;
167 }
168
169 static void efa_free_mapped(struct efa_dev *dev, void *cpu_addr,
170 dma_addr_t dma_addr,
171 size_t size, enum dma_data_direction dir)
172 {
173 dma_unmap_single(&dev->pdev->dev, dma_addr, size, dir);
174 free_pages_exact(cpu_addr, size);
175 }
176
177 int efa_query_device(struct ib_device *ibdev,
178 struct ib_device_attr *props,
179 struct ib_udata *udata)
180 {
181 struct efa_com_get_device_attr_result *dev_attr;
182 struct efa_ibv_ex_query_device_resp resp = {};
183 struct efa_dev *dev = to_edev(ibdev);
184 int err;
185
186 if (udata && udata->inlen &&
187 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
188 ibdev_dbg(ibdev,
189 "Incompatible ABI params, udata not cleared\n");
190 return -EINVAL;
191 }
192
193 dev_attr = &dev->dev_attr;
194
195 memset(props, 0, sizeof(*props));
196 props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
197 props->page_size_cap = dev_attr->page_size_cap;
198 props->vendor_id = dev->pdev->vendor;
199 props->vendor_part_id = dev->pdev->device;
200 props->hw_ver = dev->pdev->subsystem_device;
201 props->max_qp = dev_attr->max_qp;
202 props->max_cq = dev_attr->max_cq;
203 props->max_pd = dev_attr->max_pd;
204 props->max_mr = dev_attr->max_mr;
205 props->max_ah = dev_attr->max_ah;
206 props->max_cqe = dev_attr->max_cq_depth;
207 props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
208 dev_attr->max_rq_depth);
209 props->max_send_sge = dev_attr->max_sq_sge;
210 props->max_recv_sge = dev_attr->max_rq_sge;
211 props->max_sge_rd = dev_attr->max_wr_rdma_sge;
212 props->max_pkeys = 1;
213
214 if (udata && udata->outlen) {
215 resp.max_sq_sge = dev_attr->max_sq_sge;
216 resp.max_rq_sge = dev_attr->max_rq_sge;
217 resp.max_sq_wr = dev_attr->max_sq_depth;
218 resp.max_rq_wr = dev_attr->max_rq_depth;
219 resp.max_rdma_size = dev_attr->max_rdma_size;
220
221 if (is_rdma_read_cap(dev))
222 resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_READ;
223
224 err = ib_copy_to_udata(udata, &resp,
225 min(sizeof(resp), udata->outlen));
226 if (err) {
227 ibdev_dbg(ibdev,
228 "Failed to copy udata for query_device\n");
229 return err;
230 }
231 }
232
233 return 0;
234 }
235
236 int efa_query_port(struct ib_device *ibdev, u8 port,
237 struct ib_port_attr *props)
238 {
239 struct efa_dev *dev = to_edev(ibdev);
240
241 props->lmc = 1;
242
243 props->state = IB_PORT_ACTIVE;
244 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
245 props->gid_tbl_len = 1;
246 props->pkey_tbl_len = 1;
247 props->active_speed = IB_SPEED_EDR;
248 props->active_width = IB_WIDTH_4X;
249 props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
250 props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
251 props->max_msg_sz = dev->dev_attr.mtu;
252 props->max_vl_num = 1;
253
254 return 0;
255 }
256
257 int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
258 int qp_attr_mask,
259 struct ib_qp_init_attr *qp_init_attr)
260 {
261 struct efa_dev *dev = to_edev(ibqp->device);
262 struct efa_com_query_qp_params params = {};
263 struct efa_com_query_qp_result result;
264 struct efa_qp *qp = to_eqp(ibqp);
265 int err;
266
267 #define EFA_QUERY_QP_SUPP_MASK \
268 (IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
269 IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP)
270
271 if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
272 ibdev_dbg(&dev->ibdev,
273 "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
274 qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
275 return -EOPNOTSUPP;
276 }
277
278 memset(qp_attr, 0, sizeof(*qp_attr));
279 memset(qp_init_attr, 0, sizeof(*qp_init_attr));
280
281 params.qp_handle = qp->qp_handle;
282 err = efa_com_query_qp(&dev->edev, &params, &result);
283 if (err)
284 return err;
285
286 qp_attr->qp_state = result.qp_state;
287 qp_attr->qkey = result.qkey;
288 qp_attr->sq_psn = result.sq_psn;
289 qp_attr->sq_draining = result.sq_draining;
290 qp_attr->port_num = 1;
291
292 qp_attr->cap.max_send_wr = qp->max_send_wr;
293 qp_attr->cap.max_recv_wr = qp->max_recv_wr;
294 qp_attr->cap.max_send_sge = qp->max_send_sge;
295 qp_attr->cap.max_recv_sge = qp->max_recv_sge;
296 qp_attr->cap.max_inline_data = qp->max_inline_data;
297
298 qp_init_attr->qp_type = ibqp->qp_type;
299 qp_init_attr->recv_cq = ibqp->recv_cq;
300 qp_init_attr->send_cq = ibqp->send_cq;
301 qp_init_attr->qp_context = ibqp->qp_context;
302 qp_init_attr->cap = qp_attr->cap;
303
304 return 0;
305 }
306
307 int efa_query_gid(struct ib_device *ibdev, u8 port, int index,
308 union ib_gid *gid)
309 {
310 struct efa_dev *dev = to_edev(ibdev);
311
312 memcpy(gid->raw, dev->dev_attr.addr, sizeof(dev->dev_attr.addr));
313
314 return 0;
315 }
316
317 int efa_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
318 u16 *pkey)
319 {
320 if (index > 0)
321 return -EINVAL;
322
323 *pkey = 0xffff;
324 return 0;
325 }
326
327 static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
328 {
329 struct efa_com_dealloc_pd_params params = {
330 .pdn = pdn,
331 };
332
333 return efa_com_dealloc_pd(&dev->edev, &params);
334 }
335
336 int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
337 {
338 struct efa_dev *dev = to_edev(ibpd->device);
339 struct efa_ibv_alloc_pd_resp resp = {};
340 struct efa_com_alloc_pd_result result;
341 struct efa_pd *pd = to_epd(ibpd);
342 int err;
343
344 if (udata->inlen &&
345 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
346 ibdev_dbg(&dev->ibdev,
347 "Incompatible ABI params, udata not cleared\n");
348 err = -EINVAL;
349 goto err_out;
350 }
351
352 err = efa_com_alloc_pd(&dev->edev, &result);
353 if (err)
354 goto err_out;
355
356 pd->pdn = result.pdn;
357 resp.pdn = result.pdn;
358
359 if (udata->outlen) {
360 err = ib_copy_to_udata(udata, &resp,
361 min(sizeof(resp), udata->outlen));
362 if (err) {
363 ibdev_dbg(&dev->ibdev,
364 "Failed to copy udata for alloc_pd\n");
365 goto err_dealloc_pd;
366 }
367 }
368
369 ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
370
371 return 0;
372
373 err_dealloc_pd:
374 efa_pd_dealloc(dev, result.pdn);
375 err_out:
376 atomic64_inc(&dev->stats.sw_stats.alloc_pd_err);
377 return err;
378 }
379
380 void efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
381 {
382 struct efa_dev *dev = to_edev(ibpd->device);
383 struct efa_pd *pd = to_epd(ibpd);
384
385 ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
386 efa_pd_dealloc(dev, pd->pdn);
387 }
388
389 static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
390 {
391 struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
392
393 return efa_com_destroy_qp(&dev->edev, &params);
394 }
395
396 static void efa_qp_user_mmap_entries_remove(struct efa_qp *qp)
397 {
398 rdma_user_mmap_entry_remove(qp->rq_mmap_entry);
399 rdma_user_mmap_entry_remove(qp->rq_db_mmap_entry);
400 rdma_user_mmap_entry_remove(qp->llq_desc_mmap_entry);
401 rdma_user_mmap_entry_remove(qp->sq_db_mmap_entry);
402 }
403
404 int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
405 {
406 struct efa_dev *dev = to_edev(ibqp->pd->device);
407 struct efa_qp *qp = to_eqp(ibqp);
408 int err;
409
410 ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
411
412 efa_qp_user_mmap_entries_remove(qp);
413
414 err = efa_destroy_qp_handle(dev, qp->qp_handle);
415 if (err)
416 return err;
417
418 if (qp->rq_cpu_addr) {
419 ibdev_dbg(&dev->ibdev,
420 "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
421 qp->rq_cpu_addr, qp->rq_size,
422 &qp->rq_dma_addr);
423 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
424 qp->rq_size, DMA_TO_DEVICE);
425 }
426
427 kfree(qp);
428 return 0;
429 }
430
431 static struct rdma_user_mmap_entry*
432 efa_user_mmap_entry_insert(struct ib_ucontext *ucontext,
433 u64 address, size_t length,
434 u8 mmap_flag, u64 *offset)
435 {
436 struct efa_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
437 int err;
438
439 if (!entry)
440 return NULL;
441
442 entry->address = address;
443 entry->mmap_flag = mmap_flag;
444
445 err = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry,
446 length);
447 if (err) {
448 kfree(entry);
449 return NULL;
450 }
451 *offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
452
453 return &entry->rdma_entry;
454 }
455
456 static int qp_mmap_entries_setup(struct efa_qp *qp,
457 struct efa_dev *dev,
458 struct efa_ucontext *ucontext,
459 struct efa_com_create_qp_params *params,
460 struct efa_ibv_create_qp_resp *resp)
461 {
462 size_t length;
463 u64 address;
464
465 address = dev->db_bar_addr + resp->sq_db_offset;
466 qp->sq_db_mmap_entry =
467 efa_user_mmap_entry_insert(&ucontext->ibucontext,
468 address,
469 PAGE_SIZE, EFA_MMAP_IO_NC,
470 &resp->sq_db_mmap_key);
471 if (!qp->sq_db_mmap_entry)
472 return -ENOMEM;
473
474 resp->sq_db_offset &= ~PAGE_MASK;
475
476 address = dev->mem_bar_addr + resp->llq_desc_offset;
477 length = PAGE_ALIGN(params->sq_ring_size_in_bytes +
478 (resp->llq_desc_offset & ~PAGE_MASK));
479
480 qp->llq_desc_mmap_entry =
481 efa_user_mmap_entry_insert(&ucontext->ibucontext,
482 address, length,
483 EFA_MMAP_IO_WC,
484 &resp->llq_desc_mmap_key);
485 if (!qp->llq_desc_mmap_entry)
486 goto err_remove_mmap;
487
488 resp->llq_desc_offset &= ~PAGE_MASK;
489
490 if (qp->rq_size) {
491 address = dev->db_bar_addr + resp->rq_db_offset;
492
493 qp->rq_db_mmap_entry =
494 efa_user_mmap_entry_insert(&ucontext->ibucontext,
495 address, PAGE_SIZE,
496 EFA_MMAP_IO_NC,
497 &resp->rq_db_mmap_key);
498 if (!qp->rq_db_mmap_entry)
499 goto err_remove_mmap;
500
501 resp->rq_db_offset &= ~PAGE_MASK;
502
503 address = virt_to_phys(qp->rq_cpu_addr);
504 qp->rq_mmap_entry =
505 efa_user_mmap_entry_insert(&ucontext->ibucontext,
506 address, qp->rq_size,
507 EFA_MMAP_DMA_PAGE,
508 &resp->rq_mmap_key);
509 if (!qp->rq_mmap_entry)
510 goto err_remove_mmap;
511
512 resp->rq_mmap_size = qp->rq_size;
513 }
514
515 return 0;
516
517 err_remove_mmap:
518 efa_qp_user_mmap_entries_remove(qp);
519
520 return -ENOMEM;
521 }
522
523 static int efa_qp_validate_cap(struct efa_dev *dev,
524 struct ib_qp_init_attr *init_attr)
525 {
526 if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
527 ibdev_dbg(&dev->ibdev,
528 "qp: requested send wr[%u] exceeds the max[%u]\n",
529 init_attr->cap.max_send_wr,
530 dev->dev_attr.max_sq_depth);
531 return -EINVAL;
532 }
533 if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
534 ibdev_dbg(&dev->ibdev,
535 "qp: requested receive wr[%u] exceeds the max[%u]\n",
536 init_attr->cap.max_recv_wr,
537 dev->dev_attr.max_rq_depth);
538 return -EINVAL;
539 }
540 if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
541 ibdev_dbg(&dev->ibdev,
542 "qp: requested sge send[%u] exceeds the max[%u]\n",
543 init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
544 return -EINVAL;
545 }
546 if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
547 ibdev_dbg(&dev->ibdev,
548 "qp: requested sge recv[%u] exceeds the max[%u]\n",
549 init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
550 return -EINVAL;
551 }
552 if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
553 ibdev_dbg(&dev->ibdev,
554 "qp: requested inline data[%u] exceeds the max[%u]\n",
555 init_attr->cap.max_inline_data,
556 dev->dev_attr.inline_buf_size);
557 return -EINVAL;
558 }
559
560 return 0;
561 }
562
563 static int efa_qp_validate_attr(struct efa_dev *dev,
564 struct ib_qp_init_attr *init_attr)
565 {
566 if (init_attr->qp_type != IB_QPT_DRIVER &&
567 init_attr->qp_type != IB_QPT_UD) {
568 ibdev_dbg(&dev->ibdev,
569 "Unsupported qp type %d\n", init_attr->qp_type);
570 return -EOPNOTSUPP;
571 }
572
573 if (init_attr->srq) {
574 ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
575 return -EOPNOTSUPP;
576 }
577
578 if (init_attr->create_flags) {
579 ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
580 return -EOPNOTSUPP;
581 }
582
583 return 0;
584 }
585
586 struct ib_qp *efa_create_qp(struct ib_pd *ibpd,
587 struct ib_qp_init_attr *init_attr,
588 struct ib_udata *udata)
589 {
590 struct efa_com_create_qp_params create_qp_params = {};
591 struct efa_com_create_qp_result create_qp_resp;
592 struct efa_dev *dev = to_edev(ibpd->device);
593 struct efa_ibv_create_qp_resp resp = {};
594 struct efa_ibv_create_qp cmd = {};
595 struct efa_ucontext *ucontext;
596 struct efa_qp *qp;
597 int err;
598
599 ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
600 ibucontext);
601
602 err = efa_qp_validate_cap(dev, init_attr);
603 if (err)
604 goto err_out;
605
606 err = efa_qp_validate_attr(dev, init_attr);
607 if (err)
608 goto err_out;
609
610 if (offsetofend(typeof(cmd), driver_qp_type) > udata->inlen) {
611 ibdev_dbg(&dev->ibdev,
612 "Incompatible ABI params, no input udata\n");
613 err = -EINVAL;
614 goto err_out;
615 }
616
617 if (udata->inlen > sizeof(cmd) &&
618 !ib_is_udata_cleared(udata, sizeof(cmd),
619 udata->inlen - sizeof(cmd))) {
620 ibdev_dbg(&dev->ibdev,
621 "Incompatible ABI params, unknown fields in udata\n");
622 err = -EINVAL;
623 goto err_out;
624 }
625
626 err = ib_copy_from_udata(&cmd, udata,
627 min(sizeof(cmd), udata->inlen));
628 if (err) {
629 ibdev_dbg(&dev->ibdev,
630 "Cannot copy udata for create_qp\n");
631 goto err_out;
632 }
633
634 if (cmd.comp_mask) {
635 ibdev_dbg(&dev->ibdev,
636 "Incompatible ABI params, unknown fields in udata\n");
637 err = -EINVAL;
638 goto err_out;
639 }
640
641 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
642 if (!qp) {
643 err = -ENOMEM;
644 goto err_out;
645 }
646
647 create_qp_params.uarn = ucontext->uarn;
648 create_qp_params.pd = to_epd(ibpd)->pdn;
649
650 if (init_attr->qp_type == IB_QPT_UD) {
651 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
652 } else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
653 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
654 } else {
655 ibdev_dbg(&dev->ibdev,
656 "Unsupported qp type %d driver qp type %d\n",
657 init_attr->qp_type, cmd.driver_qp_type);
658 err = -EOPNOTSUPP;
659 goto err_free_qp;
660 }
661
662 ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
663 init_attr->qp_type, cmd.driver_qp_type);
664 create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
665 create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
666 create_qp_params.sq_depth = init_attr->cap.max_send_wr;
667 create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
668
669 create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
670 create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
671 qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
672 if (qp->rq_size) {
673 qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
674 qp->rq_size, DMA_TO_DEVICE);
675 if (!qp->rq_cpu_addr) {
676 err = -ENOMEM;
677 goto err_free_qp;
678 }
679
680 ibdev_dbg(&dev->ibdev,
681 "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
682 qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
683 create_qp_params.rq_base_addr = qp->rq_dma_addr;
684 }
685
686 err = efa_com_create_qp(&dev->edev, &create_qp_params,
687 &create_qp_resp);
688 if (err)
689 goto err_free_mapped;
690
691 resp.sq_db_offset = create_qp_resp.sq_db_offset;
692 resp.rq_db_offset = create_qp_resp.rq_db_offset;
693 resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
694 resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
695 resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
696
697 err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
698 &resp);
699 if (err)
700 goto err_destroy_qp;
701
702 qp->qp_handle = create_qp_resp.qp_handle;
703 qp->ibqp.qp_num = create_qp_resp.qp_num;
704 qp->ibqp.qp_type = init_attr->qp_type;
705 qp->max_send_wr = init_attr->cap.max_send_wr;
706 qp->max_recv_wr = init_attr->cap.max_recv_wr;
707 qp->max_send_sge = init_attr->cap.max_send_sge;
708 qp->max_recv_sge = init_attr->cap.max_recv_sge;
709 qp->max_inline_data = init_attr->cap.max_inline_data;
710
711 if (udata->outlen) {
712 err = ib_copy_to_udata(udata, &resp,
713 min(sizeof(resp), udata->outlen));
714 if (err) {
715 ibdev_dbg(&dev->ibdev,
716 "Failed to copy udata for qp[%u]\n",
717 create_qp_resp.qp_num);
718 goto err_remove_mmap_entries;
719 }
720 }
721
722 ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
723
724 return &qp->ibqp;
725
726 err_remove_mmap_entries:
727 efa_qp_user_mmap_entries_remove(qp);
728 err_destroy_qp:
729 efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
730 err_free_mapped:
731 if (qp->rq_size)
732 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
733 qp->rq_size, DMA_TO_DEVICE);
734 err_free_qp:
735 kfree(qp);
736 err_out:
737 atomic64_inc(&dev->stats.sw_stats.create_qp_err);
738 return ERR_PTR(err);
739 }
740
741 static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
742 struct ib_qp_attr *qp_attr, int qp_attr_mask,
743 enum ib_qp_state cur_state,
744 enum ib_qp_state new_state)
745 {
746 #define EFA_MODIFY_QP_SUPP_MASK \
747 (IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
748 IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN)
749
750 if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
751 ibdev_dbg(&dev->ibdev,
752 "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
753 qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
754 return -EOPNOTSUPP;
755 }
756
757 if (!ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
758 qp_attr_mask)) {
759 ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
760 return -EINVAL;
761 }
762
763 if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
764 ibdev_dbg(&dev->ibdev, "Can't change port num\n");
765 return -EOPNOTSUPP;
766 }
767
768 if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
769 ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
770 return -EOPNOTSUPP;
771 }
772
773 return 0;
774 }
775
776 int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
777 int qp_attr_mask, struct ib_udata *udata)
778 {
779 struct efa_dev *dev = to_edev(ibqp->device);
780 struct efa_com_modify_qp_params params = {};
781 struct efa_qp *qp = to_eqp(ibqp);
782 enum ib_qp_state cur_state;
783 enum ib_qp_state new_state;
784 int err;
785
786 if (udata->inlen &&
787 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
788 ibdev_dbg(&dev->ibdev,
789 "Incompatible ABI params, udata not cleared\n");
790 return -EINVAL;
791 }
792
793 cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
794 qp->state;
795 new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
796
797 err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
798 new_state);
799 if (err)
800 return err;
801
802 params.qp_handle = qp->qp_handle;
803
804 if (qp_attr_mask & IB_QP_STATE) {
805 params.modify_mask |= BIT(EFA_ADMIN_QP_STATE_BIT) |
806 BIT(EFA_ADMIN_CUR_QP_STATE_BIT);
807 params.cur_qp_state = qp_attr->cur_qp_state;
808 params.qp_state = qp_attr->qp_state;
809 }
810
811 if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
812 params.modify_mask |=
813 BIT(EFA_ADMIN_SQ_DRAINED_ASYNC_NOTIFY_BIT);
814 params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
815 }
816
817 if (qp_attr_mask & IB_QP_QKEY) {
818 params.modify_mask |= BIT(EFA_ADMIN_QKEY_BIT);
819 params.qkey = qp_attr->qkey;
820 }
821
822 if (qp_attr_mask & IB_QP_SQ_PSN) {
823 params.modify_mask |= BIT(EFA_ADMIN_SQ_PSN_BIT);
824 params.sq_psn = qp_attr->sq_psn;
825 }
826
827 err = efa_com_modify_qp(&dev->edev, &params);
828 if (err)
829 return err;
830
831 qp->state = new_state;
832
833 return 0;
834 }
835
836 static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
837 {
838 struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
839
840 return efa_com_destroy_cq(&dev->edev, &params);
841 }
842
843 void efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
844 {
845 struct efa_dev *dev = to_edev(ibcq->device);
846 struct efa_cq *cq = to_ecq(ibcq);
847
848 ibdev_dbg(&dev->ibdev,
849 "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
850 cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
851
852 rdma_user_mmap_entry_remove(cq->mmap_entry);
853 efa_destroy_cq_idx(dev, cq->cq_idx);
854 efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
855 DMA_FROM_DEVICE);
856 }
857
858 static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
859 struct efa_ibv_create_cq_resp *resp)
860 {
861 resp->q_mmap_size = cq->size;
862 cq->mmap_entry = efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
863 virt_to_phys(cq->cpu_addr),
864 cq->size, EFA_MMAP_DMA_PAGE,
865 &resp->q_mmap_key);
866 if (!cq->mmap_entry)
867 return -ENOMEM;
868
869 return 0;
870 }
871
872 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
873 struct ib_udata *udata)
874 {
875 struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
876 udata, struct efa_ucontext, ibucontext);
877 struct efa_ibv_create_cq_resp resp = {};
878 struct efa_com_create_cq_params params;
879 struct efa_com_create_cq_result result;
880 struct ib_device *ibdev = ibcq->device;
881 struct efa_dev *dev = to_edev(ibdev);
882 struct efa_ibv_create_cq cmd = {};
883 struct efa_cq *cq = to_ecq(ibcq);
884 int entries = attr->cqe;
885 int err;
886
887 ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
888
889 if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
890 ibdev_dbg(ibdev,
891 "cq: requested entries[%u] non-positive or greater than max[%u]\n",
892 entries, dev->dev_attr.max_cq_depth);
893 err = -EINVAL;
894 goto err_out;
895 }
896
897 if (offsetofend(typeof(cmd), num_sub_cqs) > udata->inlen) {
898 ibdev_dbg(ibdev,
899 "Incompatible ABI params, no input udata\n");
900 err = -EINVAL;
901 goto err_out;
902 }
903
904 if (udata->inlen > sizeof(cmd) &&
905 !ib_is_udata_cleared(udata, sizeof(cmd),
906 udata->inlen - sizeof(cmd))) {
907 ibdev_dbg(ibdev,
908 "Incompatible ABI params, unknown fields in udata\n");
909 err = -EINVAL;
910 goto err_out;
911 }
912
913 err = ib_copy_from_udata(&cmd, udata,
914 min(sizeof(cmd), udata->inlen));
915 if (err) {
916 ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
917 goto err_out;
918 }
919
920 if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_50)) {
921 ibdev_dbg(ibdev,
922 "Incompatible ABI params, unknown fields in udata\n");
923 err = -EINVAL;
924 goto err_out;
925 }
926
927 if (!cmd.cq_entry_size) {
928 ibdev_dbg(ibdev,
929 "Invalid entry size [%u]\n", cmd.cq_entry_size);
930 err = -EINVAL;
931 goto err_out;
932 }
933
934 if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
935 ibdev_dbg(ibdev,
936 "Invalid number of sub cqs[%u] expected[%u]\n",
937 cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
938 err = -EINVAL;
939 goto err_out;
940 }
941
942 cq->ucontext = ucontext;
943 cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
944 cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
945 DMA_FROM_DEVICE);
946 if (!cq->cpu_addr) {
947 err = -ENOMEM;
948 goto err_out;
949 }
950
951 params.uarn = cq->ucontext->uarn;
952 params.cq_depth = entries;
953 params.dma_addr = cq->dma_addr;
954 params.entry_size_in_bytes = cmd.cq_entry_size;
955 params.num_sub_cqs = cmd.num_sub_cqs;
956 err = efa_com_create_cq(&dev->edev, &params, &result);
957 if (err)
958 goto err_free_mapped;
959
960 resp.cq_idx = result.cq_idx;
961 cq->cq_idx = result.cq_idx;
962 cq->ibcq.cqe = result.actual_depth;
963 WARN_ON_ONCE(entries != result.actual_depth);
964
965 err = cq_mmap_entries_setup(dev, cq, &resp);
966 if (err) {
967 ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n",
968 cq->cq_idx);
969 goto err_destroy_cq;
970 }
971
972 if (udata->outlen) {
973 err = ib_copy_to_udata(udata, &resp,
974 min(sizeof(resp), udata->outlen));
975 if (err) {
976 ibdev_dbg(ibdev,
977 "Failed to copy udata for create_cq\n");
978 goto err_remove_mmap;
979 }
980 }
981
982 ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
983 cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
984
985 return 0;
986
987 err_remove_mmap:
988 rdma_user_mmap_entry_remove(cq->mmap_entry);
989 err_destroy_cq:
990 efa_destroy_cq_idx(dev, cq->cq_idx);
991 err_free_mapped:
992 efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
993 DMA_FROM_DEVICE);
994
995 err_out:
996 atomic64_inc(&dev->stats.sw_stats.create_cq_err);
997 return err;
998 }
999
1000 static int umem_to_page_list(struct efa_dev *dev,
1001 struct ib_umem *umem,
1002 u64 *page_list,
1003 u32 hp_cnt,
1004 u8 hp_shift)
1005 {
1006 u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT);
1007 struct ib_block_iter biter;
1008 unsigned int hp_idx = 0;
1009
1010 ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n",
1011 hp_cnt, pages_in_hp);
1012
1013 rdma_for_each_block(umem->sg_head.sgl, &biter, umem->nmap,
1014 BIT(hp_shift))
1015 page_list[hp_idx++] = rdma_block_iter_dma_address(&biter);
1016
1017 return 0;
1018 }
1019
1020 static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1021 {
1022 struct scatterlist *sglist;
1023 struct page *pg;
1024 int i;
1025
1026 sglist = kcalloc(page_cnt, sizeof(*sglist), GFP_KERNEL);
1027 if (!sglist)
1028 return NULL;
1029 sg_init_table(sglist, page_cnt);
1030 for (i = 0; i < page_cnt; i++) {
1031 pg = vmalloc_to_page(buf);
1032 if (!pg)
1033 goto err;
1034 sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1035 buf += PAGE_SIZE / sizeof(*buf);
1036 }
1037 return sglist;
1038
1039 err:
1040 kfree(sglist);
1041 return NULL;
1042 }
1043
1044 /*
1045 * create a chunk list of physical pages dma addresses from the supplied
1046 * scatter gather list
1047 */
1048 static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1049 {
1050 struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1051 int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1052 struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1053 unsigned int chunk_list_size, chunk_idx, payload_idx;
1054 int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1055 struct efa_com_ctrl_buff_info *ctrl_buf;
1056 u64 *cur_chunk_buf, *prev_chunk_buf;
1057 struct ib_block_iter biter;
1058 dma_addr_t dma_addr;
1059 int i;
1060
1061 /* allocate a chunk list that consists of 4KB chunks */
1062 chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1063
1064 chunk_list->size = chunk_list_size;
1065 chunk_list->chunks = kcalloc(chunk_list_size,
1066 sizeof(*chunk_list->chunks),
1067 GFP_KERNEL);
1068 if (!chunk_list->chunks)
1069 return -ENOMEM;
1070
1071 ibdev_dbg(&dev->ibdev,
1072 "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1073 page_cnt);
1074
1075 /* allocate chunk buffers: */
1076 for (i = 0; i < chunk_list_size; i++) {
1077 chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1078 if (!chunk_list->chunks[i].buf)
1079 goto chunk_list_dealloc;
1080
1081 chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1082 }
1083 chunk_list->chunks[chunk_list_size - 1].length =
1084 ((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1085 EFA_CHUNK_PTR_SIZE;
1086
1087 /* fill the dma addresses of sg list pages to chunks: */
1088 chunk_idx = 0;
1089 payload_idx = 0;
1090 cur_chunk_buf = chunk_list->chunks[0].buf;
1091 rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt,
1092 EFA_CHUNK_PAYLOAD_SIZE) {
1093 cur_chunk_buf[payload_idx++] =
1094 rdma_block_iter_dma_address(&biter);
1095
1096 if (payload_idx == EFA_PTRS_PER_CHUNK) {
1097 chunk_idx++;
1098 cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1099 payload_idx = 0;
1100 }
1101 }
1102
1103 /* map chunks to dma and fill chunks next ptrs */
1104 for (i = chunk_list_size - 1; i >= 0; i--) {
1105 dma_addr = dma_map_single(&dev->pdev->dev,
1106 chunk_list->chunks[i].buf,
1107 chunk_list->chunks[i].length,
1108 DMA_TO_DEVICE);
1109 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1110 ibdev_err(&dev->ibdev,
1111 "chunk[%u] dma_map_failed\n", i);
1112 goto chunk_list_unmap;
1113 }
1114
1115 chunk_list->chunks[i].dma_addr = dma_addr;
1116 ibdev_dbg(&dev->ibdev,
1117 "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1118
1119 if (!i)
1120 break;
1121
1122 prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1123
1124 ctrl_buf = (struct efa_com_ctrl_buff_info *)
1125 &prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1126 ctrl_buf->length = chunk_list->chunks[i].length;
1127
1128 efa_com_set_dma_addr(dma_addr,
1129 &ctrl_buf->address.mem_addr_high,
1130 &ctrl_buf->address.mem_addr_low);
1131 }
1132
1133 return 0;
1134
1135 chunk_list_unmap:
1136 for (; i < chunk_list_size; i++) {
1137 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1138 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1139 }
1140 chunk_list_dealloc:
1141 for (i = 0; i < chunk_list_size; i++)
1142 kfree(chunk_list->chunks[i].buf);
1143
1144 kfree(chunk_list->chunks);
1145 return -ENOMEM;
1146 }
1147
1148 static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1149 {
1150 struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1151 int i;
1152
1153 for (i = 0; i < chunk_list->size; i++) {
1154 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1155 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1156 kfree(chunk_list->chunks[i].buf);
1157 }
1158
1159 kfree(chunk_list->chunks);
1160 }
1161
1162 /* initialize pbl continuous mode: map pbl buffer to a dma address. */
1163 static int pbl_continuous_initialize(struct efa_dev *dev,
1164 struct pbl_context *pbl)
1165 {
1166 dma_addr_t dma_addr;
1167
1168 dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1169 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1170 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1171 ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1172 return -ENOMEM;
1173 }
1174
1175 pbl->phys.continuous.dma_addr = dma_addr;
1176 ibdev_dbg(&dev->ibdev,
1177 "pbl continuous - dma_addr = %pad, size[%u]\n",
1178 &dma_addr, pbl->pbl_buf_size_in_bytes);
1179
1180 return 0;
1181 }
1182
1183 /*
1184 * initialize pbl indirect mode:
1185 * create a chunk list out of the dma addresses of the physical pages of
1186 * pbl buffer.
1187 */
1188 static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1189 {
1190 u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, PAGE_SIZE);
1191 struct scatterlist *sgl;
1192 int sg_dma_cnt, err;
1193
1194 BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1195 sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1196 if (!sgl)
1197 return -ENOMEM;
1198
1199 sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1200 if (!sg_dma_cnt) {
1201 err = -EINVAL;
1202 goto err_map;
1203 }
1204
1205 pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1206 pbl->phys.indirect.sgl = sgl;
1207 pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1208 err = pbl_chunk_list_create(dev, pbl);
1209 if (err) {
1210 ibdev_dbg(&dev->ibdev,
1211 "chunk_list creation failed[%d]\n", err);
1212 goto err_chunk;
1213 }
1214
1215 ibdev_dbg(&dev->ibdev,
1216 "pbl indirect - size[%u], chunks[%u]\n",
1217 pbl->pbl_buf_size_in_bytes,
1218 pbl->phys.indirect.chunk_list.size);
1219
1220 return 0;
1221
1222 err_chunk:
1223 dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1224 err_map:
1225 kfree(sgl);
1226 return err;
1227 }
1228
1229 static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1230 {
1231 pbl_chunk_list_destroy(dev, pbl);
1232 dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1233 pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1234 kfree(pbl->phys.indirect.sgl);
1235 }
1236
1237 /* create a page buffer list from a mapped user memory region */
1238 static int pbl_create(struct efa_dev *dev,
1239 struct pbl_context *pbl,
1240 struct ib_umem *umem,
1241 int hp_cnt,
1242 u8 hp_shift)
1243 {
1244 int err;
1245
1246 pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1247 pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL);
1248 if (!pbl->pbl_buf)
1249 return -ENOMEM;
1250
1251 if (is_vmalloc_addr(pbl->pbl_buf)) {
1252 pbl->physically_continuous = 0;
1253 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1254 hp_shift);
1255 if (err)
1256 goto err_free;
1257
1258 err = pbl_indirect_initialize(dev, pbl);
1259 if (err)
1260 goto err_free;
1261 } else {
1262 pbl->physically_continuous = 1;
1263 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1264 hp_shift);
1265 if (err)
1266 goto err_free;
1267
1268 err = pbl_continuous_initialize(dev, pbl);
1269 if (err)
1270 goto err_free;
1271 }
1272
1273 ibdev_dbg(&dev->ibdev,
1274 "user_pbl_created: user_pages[%u], continuous[%u]\n",
1275 hp_cnt, pbl->physically_continuous);
1276
1277 return 0;
1278
1279 err_free:
1280 kvfree(pbl->pbl_buf);
1281 return err;
1282 }
1283
1284 static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1285 {
1286 if (pbl->physically_continuous)
1287 dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1288 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1289 else
1290 pbl_indirect_terminate(dev, pbl);
1291
1292 kvfree(pbl->pbl_buf);
1293 }
1294
1295 static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1296 struct efa_com_reg_mr_params *params)
1297 {
1298 int err;
1299
1300 params->inline_pbl = 1;
1301 err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1302 params->page_num, params->page_shift);
1303 if (err)
1304 return err;
1305
1306 ibdev_dbg(&dev->ibdev,
1307 "inline_pbl_array - pages[%u]\n", params->page_num);
1308
1309 return 0;
1310 }
1311
1312 static int efa_create_pbl(struct efa_dev *dev,
1313 struct pbl_context *pbl,
1314 struct efa_mr *mr,
1315 struct efa_com_reg_mr_params *params)
1316 {
1317 int err;
1318
1319 err = pbl_create(dev, pbl, mr->umem, params->page_num,
1320 params->page_shift);
1321 if (err) {
1322 ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1323 return err;
1324 }
1325
1326 params->inline_pbl = 0;
1327 params->indirect = !pbl->physically_continuous;
1328 if (pbl->physically_continuous) {
1329 params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1330
1331 efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1332 &params->pbl.pbl.address.mem_addr_high,
1333 &params->pbl.pbl.address.mem_addr_low);
1334 } else {
1335 params->pbl.pbl.length =
1336 pbl->phys.indirect.chunk_list.chunks[0].length;
1337
1338 efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1339 &params->pbl.pbl.address.mem_addr_high,
1340 &params->pbl.pbl.address.mem_addr_low);
1341 }
1342
1343 return 0;
1344 }
1345
1346 struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1347 u64 virt_addr, int access_flags,
1348 struct ib_udata *udata)
1349 {
1350 struct efa_dev *dev = to_edev(ibpd->device);
1351 struct efa_com_reg_mr_params params = {};
1352 struct efa_com_reg_mr_result result = {};
1353 struct pbl_context pbl;
1354 int supp_access_flags;
1355 unsigned int pg_sz;
1356 struct efa_mr *mr;
1357 int inline_size;
1358 int err;
1359
1360 if (udata && udata->inlen &&
1361 !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
1362 ibdev_dbg(&dev->ibdev,
1363 "Incompatible ABI params, udata not cleared\n");
1364 err = -EINVAL;
1365 goto err_out;
1366 }
1367
1368 supp_access_flags =
1369 IB_ACCESS_LOCAL_WRITE |
1370 (is_rdma_read_cap(dev) ? IB_ACCESS_REMOTE_READ : 0);
1371
1372 access_flags &= ~IB_ACCESS_OPTIONAL;
1373 if (access_flags & ~supp_access_flags) {
1374 ibdev_dbg(&dev->ibdev,
1375 "Unsupported access flags[%#x], supported[%#x]\n",
1376 access_flags, supp_access_flags);
1377 err = -EOPNOTSUPP;
1378 goto err_out;
1379 }
1380
1381 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1382 if (!mr) {
1383 err = -ENOMEM;
1384 goto err_out;
1385 }
1386
1387 mr->umem = ib_umem_get(ibpd->device, start, length, access_flags);
1388 if (IS_ERR(mr->umem)) {
1389 err = PTR_ERR(mr->umem);
1390 ibdev_dbg(&dev->ibdev,
1391 "Failed to pin and map user space memory[%d]\n", err);
1392 goto err_free;
1393 }
1394
1395 params.pd = to_epd(ibpd)->pdn;
1396 params.iova = virt_addr;
1397 params.mr_length_in_bytes = length;
1398 params.permissions = access_flags;
1399
1400 pg_sz = ib_umem_find_best_pgsz(mr->umem,
1401 dev->dev_attr.page_size_cap,
1402 virt_addr);
1403 if (!pg_sz) {
1404 err = -EOPNOTSUPP;
1405 ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n",
1406 dev->dev_attr.page_size_cap);
1407 goto err_unmap;
1408 }
1409
1410 params.page_shift = __ffs(pg_sz);
1411 params.page_num = DIV_ROUND_UP(length + (start & (pg_sz - 1)),
1412 pg_sz);
1413
1414 ibdev_dbg(&dev->ibdev,
1415 "start %#llx length %#llx params.page_shift %u params.page_num %u\n",
1416 start, length, params.page_shift, params.page_num);
1417
1418 inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1419 if (params.page_num <= inline_size) {
1420 err = efa_create_inline_pbl(dev, mr, &params);
1421 if (err)
1422 goto err_unmap;
1423
1424 err = efa_com_register_mr(&dev->edev, &params, &result);
1425 if (err)
1426 goto err_unmap;
1427 } else {
1428 err = efa_create_pbl(dev, &pbl, mr, &params);
1429 if (err)
1430 goto err_unmap;
1431
1432 err = efa_com_register_mr(&dev->edev, &params, &result);
1433 pbl_destroy(dev, &pbl);
1434
1435 if (err)
1436 goto err_unmap;
1437 }
1438
1439 mr->ibmr.lkey = result.l_key;
1440 mr->ibmr.rkey = result.r_key;
1441 mr->ibmr.length = length;
1442 ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1443
1444 return &mr->ibmr;
1445
1446 err_unmap:
1447 ib_umem_release(mr->umem);
1448 err_free:
1449 kfree(mr);
1450 err_out:
1451 atomic64_inc(&dev->stats.sw_stats.reg_mr_err);
1452 return ERR_PTR(err);
1453 }
1454
1455 int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1456 {
1457 struct efa_dev *dev = to_edev(ibmr->device);
1458 struct efa_com_dereg_mr_params params;
1459 struct efa_mr *mr = to_emr(ibmr);
1460 int err;
1461
1462 ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1463
1464 params.l_key = mr->ibmr.lkey;
1465 err = efa_com_dereg_mr(&dev->edev, &params);
1466 if (err)
1467 return err;
1468
1469 ib_umem_release(mr->umem);
1470 kfree(mr);
1471
1472 return 0;
1473 }
1474
1475 int efa_get_port_immutable(struct ib_device *ibdev, u8 port_num,
1476 struct ib_port_immutable *immutable)
1477 {
1478 struct ib_port_attr attr;
1479 int err;
1480
1481 err = ib_query_port(ibdev, port_num, &attr);
1482 if (err) {
1483 ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1484 return err;
1485 }
1486
1487 immutable->pkey_tbl_len = attr.pkey_tbl_len;
1488 immutable->gid_tbl_len = attr.gid_tbl_len;
1489
1490 return 0;
1491 }
1492
1493 static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1494 {
1495 struct efa_com_dealloc_uar_params params = {
1496 .uarn = uarn,
1497 };
1498
1499 return efa_com_dealloc_uar(&dev->edev, &params);
1500 }
1501
1502 int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1503 {
1504 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1505 struct efa_dev *dev = to_edev(ibucontext->device);
1506 struct efa_ibv_alloc_ucontext_resp resp = {};
1507 struct efa_com_alloc_uar_result result;
1508 int err;
1509
1510 /*
1511 * it's fine if the driver does not know all request fields,
1512 * we will ack input fields in our response.
1513 */
1514
1515 err = efa_com_alloc_uar(&dev->edev, &result);
1516 if (err)
1517 goto err_out;
1518
1519 ucontext->uarn = result.uarn;
1520
1521 resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1522 resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1523 resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1524 resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1525 resp.max_llq_size = dev->dev_attr.max_llq_size;
1526
1527 if (udata && udata->outlen) {
1528 err = ib_copy_to_udata(udata, &resp,
1529 min(sizeof(resp), udata->outlen));
1530 if (err)
1531 goto err_dealloc_uar;
1532 }
1533
1534 return 0;
1535
1536 err_dealloc_uar:
1537 efa_dealloc_uar(dev, result.uarn);
1538 err_out:
1539 atomic64_inc(&dev->stats.sw_stats.alloc_ucontext_err);
1540 return err;
1541 }
1542
1543 void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
1544 {
1545 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1546 struct efa_dev *dev = to_edev(ibucontext->device);
1547
1548 efa_dealloc_uar(dev, ucontext->uarn);
1549 }
1550
1551 void efa_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
1552 {
1553 struct efa_user_mmap_entry *entry = to_emmap(rdma_entry);
1554
1555 kfree(entry);
1556 }
1557
1558 static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
1559 struct vm_area_struct *vma)
1560 {
1561 struct rdma_user_mmap_entry *rdma_entry;
1562 struct efa_user_mmap_entry *entry;
1563 unsigned long va;
1564 int err = 0;
1565 u64 pfn;
1566
1567 rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
1568 if (!rdma_entry) {
1569 ibdev_dbg(&dev->ibdev,
1570 "pgoff[%#lx] does not have valid entry\n",
1571 vma->vm_pgoff);
1572 return -EINVAL;
1573 }
1574 entry = to_emmap(rdma_entry);
1575
1576 ibdev_dbg(&dev->ibdev,
1577 "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
1578 entry->address, rdma_entry->npages * PAGE_SIZE,
1579 entry->mmap_flag);
1580
1581 pfn = entry->address >> PAGE_SHIFT;
1582 switch (entry->mmap_flag) {
1583 case EFA_MMAP_IO_NC:
1584 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1585 entry->rdma_entry.npages * PAGE_SIZE,
1586 pgprot_noncached(vma->vm_page_prot),
1587 rdma_entry);
1588 break;
1589 case EFA_MMAP_IO_WC:
1590 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1591 entry->rdma_entry.npages * PAGE_SIZE,
1592 pgprot_writecombine(vma->vm_page_prot),
1593 rdma_entry);
1594 break;
1595 case EFA_MMAP_DMA_PAGE:
1596 for (va = vma->vm_start; va < vma->vm_end;
1597 va += PAGE_SIZE, pfn++) {
1598 err = vm_insert_page(vma, va, pfn_to_page(pfn));
1599 if (err)
1600 break;
1601 }
1602 break;
1603 default:
1604 err = -EINVAL;
1605 }
1606
1607 if (err)
1608 ibdev_dbg(
1609 &dev->ibdev,
1610 "Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
1611 entry->address, rdma_entry->npages * PAGE_SIZE,
1612 entry->mmap_flag, err);
1613
1614 rdma_user_mmap_entry_put(rdma_entry);
1615 return err;
1616 }
1617
1618 int efa_mmap(struct ib_ucontext *ibucontext,
1619 struct vm_area_struct *vma)
1620 {
1621 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1622 struct efa_dev *dev = to_edev(ibucontext->device);
1623 size_t length = vma->vm_end - vma->vm_start;
1624
1625 ibdev_dbg(&dev->ibdev,
1626 "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
1627 vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
1628
1629 return __efa_mmap(dev, ucontext, vma);
1630 }
1631
1632 static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
1633 {
1634 struct efa_com_destroy_ah_params params = {
1635 .ah = ah->ah,
1636 .pdn = to_epd(ah->ibah.pd)->pdn,
1637 };
1638
1639 return efa_com_destroy_ah(&dev->edev, &params);
1640 }
1641
1642 int efa_create_ah(struct ib_ah *ibah,
1643 struct rdma_ah_attr *ah_attr,
1644 u32 flags,
1645 struct ib_udata *udata)
1646 {
1647 struct efa_dev *dev = to_edev(ibah->device);
1648 struct efa_com_create_ah_params params = {};
1649 struct efa_ibv_create_ah_resp resp = {};
1650 struct efa_com_create_ah_result result;
1651 struct efa_ah *ah = to_eah(ibah);
1652 int err;
1653
1654 if (!(flags & RDMA_CREATE_AH_SLEEPABLE)) {
1655 ibdev_dbg(&dev->ibdev,
1656 "Create address handle is not supported in atomic context\n");
1657 err = -EOPNOTSUPP;
1658 goto err_out;
1659 }
1660
1661 if (udata->inlen &&
1662 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1663 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1664 err = -EINVAL;
1665 goto err_out;
1666 }
1667
1668 memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
1669 sizeof(params.dest_addr));
1670 params.pdn = to_epd(ibah->pd)->pdn;
1671 err = efa_com_create_ah(&dev->edev, &params, &result);
1672 if (err)
1673 goto err_out;
1674
1675 memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
1676 ah->ah = result.ah;
1677
1678 resp.efa_address_handle = result.ah;
1679
1680 if (udata->outlen) {
1681 err = ib_copy_to_udata(udata, &resp,
1682 min(sizeof(resp), udata->outlen));
1683 if (err) {
1684 ibdev_dbg(&dev->ibdev,
1685 "Failed to copy udata for create_ah response\n");
1686 goto err_destroy_ah;
1687 }
1688 }
1689 ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
1690
1691 return 0;
1692
1693 err_destroy_ah:
1694 efa_ah_destroy(dev, ah);
1695 err_out:
1696 atomic64_inc(&dev->stats.sw_stats.create_ah_err);
1697 return err;
1698 }
1699
1700 void efa_destroy_ah(struct ib_ah *ibah, u32 flags)
1701 {
1702 struct efa_dev *dev = to_edev(ibah->pd->device);
1703 struct efa_ah *ah = to_eah(ibah);
1704
1705 ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
1706
1707 if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
1708 ibdev_dbg(&dev->ibdev,
1709 "Destroy address handle is not supported in atomic context\n");
1710 return;
1711 }
1712
1713 efa_ah_destroy(dev, ah);
1714 }
1715
1716 struct rdma_hw_stats *efa_alloc_hw_stats(struct ib_device *ibdev, u8 port_num)
1717 {
1718 return rdma_alloc_hw_stats_struct(efa_stats_names,
1719 ARRAY_SIZE(efa_stats_names),
1720 RDMA_HW_STATS_DEFAULT_LIFESPAN);
1721 }
1722
1723 int efa_get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1724 u8 port_num, int index)
1725 {
1726 struct efa_com_get_stats_params params = {};
1727 union efa_com_get_stats_result result;
1728 struct efa_dev *dev = to_edev(ibdev);
1729 struct efa_com_basic_stats *bs;
1730 struct efa_com_stats_admin *as;
1731 struct efa_stats *s;
1732 int err;
1733
1734 params.type = EFA_ADMIN_GET_STATS_TYPE_BASIC;
1735 params.scope = EFA_ADMIN_GET_STATS_SCOPE_ALL;
1736
1737 err = efa_com_get_stats(&dev->edev, &params, &result);
1738 if (err)
1739 return err;
1740
1741 bs = &result.basic_stats;
1742 stats->value[EFA_TX_BYTES] = bs->tx_bytes;
1743 stats->value[EFA_TX_PKTS] = bs->tx_pkts;
1744 stats->value[EFA_RX_BYTES] = bs->rx_bytes;
1745 stats->value[EFA_RX_PKTS] = bs->rx_pkts;
1746 stats->value[EFA_RX_DROPS] = bs->rx_drops;
1747
1748 as = &dev->edev.aq.stats;
1749 stats->value[EFA_SUBMITTED_CMDS] = atomic64_read(&as->submitted_cmd);
1750 stats->value[EFA_COMPLETED_CMDS] = atomic64_read(&as->completed_cmd);
1751 stats->value[EFA_NO_COMPLETION_CMDS] = atomic64_read(&as->no_completion);
1752
1753 s = &dev->stats;
1754 stats->value[EFA_KEEP_ALIVE_RCVD] = atomic64_read(&s->keep_alive_rcvd);
1755 stats->value[EFA_ALLOC_PD_ERR] = atomic64_read(&s->sw_stats.alloc_pd_err);
1756 stats->value[EFA_CREATE_QP_ERR] = atomic64_read(&s->sw_stats.create_qp_err);
1757 stats->value[EFA_REG_MR_ERR] = atomic64_read(&s->sw_stats.reg_mr_err);
1758 stats->value[EFA_ALLOC_UCONTEXT_ERR] = atomic64_read(&s->sw_stats.alloc_ucontext_err);
1759 stats->value[EFA_CREATE_AH_ERR] = atomic64_read(&s->sw_stats.create_ah_err);
1760
1761 return ARRAY_SIZE(efa_stats_names);
1762 }
1763
1764 enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
1765 u8 port_num)
1766 {
1767 return IB_LINK_LAYER_UNSPECIFIED;
1768 }
1769
Linux with added FPC-III support