1/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
2/*
3 * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
4 * Copyright (c) 2020, Intel Corporation. All rights reserved.
5 */
6
7#ifndef MLX5_IB_H
8#define MLX5_IB_H
9
10#include <linux/kernel.h>
11#include <linux/sched.h>
12#include <rdma/ib_verbs.h>
13#include <rdma/ib_umem.h>
14#include <rdma/ib_smi.h>
15#include <linux/mlx5/driver.h>
16#include <linux/mlx5/cq.h>
17#include <linux/mlx5/fs.h>
18#include <linux/mlx5/qp.h>
19#include <linux/types.h>
20#include <linux/mlx5/transobj.h>
21#include <rdma/ib_user_verbs.h>
22#include <rdma/mlx5-abi.h>
23#include <rdma/uverbs_ioctl.h>
24#include <rdma/mlx5_user_ioctl_cmds.h>
25#include <rdma/mlx5_user_ioctl_verbs.h>
26
27#include "srq.h"
28#include "qp.h"
29#include "macsec.h"
30
31#define mlx5_ib_dbg(_dev, format, arg...) \
32 dev_dbg(&(_dev)->ib_dev.dev, "%s:%d:(pid %d): " format, __func__, \
33 __LINE__, current->pid, ##arg)
34
35#define mlx5_ib_err(_dev, format, arg...) \
36 dev_err(&(_dev)->ib_dev.dev, "%s:%d:(pid %d): " format, __func__, \
37 __LINE__, current->pid, ##arg)
38
39#define mlx5_ib_warn(_dev, format, arg...) \
40 dev_warn(&(_dev)->ib_dev.dev, "%s:%d:(pid %d): " format, __func__, \
41 __LINE__, current->pid, ##arg)
42
43#define mlx5_ib_log(lvl, _dev, format, arg...) \
44 dev_printk(lvl, &(_dev)->ib_dev.dev, "%s:%d:(pid %d): " format, \
45 __func__, __LINE__, current->pid, ##arg)
46
47#define MLX5_IB_DEFAULT_UIDX 0xffffff
48#define MLX5_USER_ASSIGNED_UIDX_MASK __mlx5_mask(qpc, user_index)
49
50static __always_inline unsigned long
51__mlx5_log_page_size_to_bitmap(unsigned int log_pgsz_bits,
52 unsigned int pgsz_shift)
53{
54 unsigned int largest_pg_shift =
55 min_t(unsigned long, (1ULL << log_pgsz_bits) - 1 + pgsz_shift,
56 BITS_PER_LONG - 1);
57
58 /*
59 * Despite a command allowing it, the device does not support lower than
60 * 4k page size.
61 */
62 pgsz_shift = max_t(unsigned int, MLX5_ADAPTER_PAGE_SHIFT, pgsz_shift);
63 return GENMASK(largest_pg_shift, pgsz_shift);
64}
65
66static __always_inline unsigned long
67__mlx5_page_offset_to_bitmask(unsigned int page_offset_bits,
68 unsigned int offset_shift)
69{
70 unsigned int largest_offset_shift =
71 min_t(unsigned long, page_offset_bits - 1 + offset_shift,
72 BITS_PER_LONG - 1);
73
74 return GENMASK(largest_offset_shift, offset_shift);
75}
76
77/*
78 * QP/CQ/WQ/etc type commands take a page offset that satisifies:
79 * page_offset_quantized * (page_size/scale) = page_offset
80 * Which restricts allowed page sizes to ones that satisify the above.
81 */
82unsigned long __mlx5_umem_find_best_quantized_pgoff(
83 struct ib_umem *umem, unsigned long pgsz_bitmap,
84 unsigned int page_offset_bits, u64 pgoff_bitmask, unsigned int scale,
85 unsigned int *page_offset_quantized);
86#define mlx5_umem_find_best_quantized_pgoff(umem, typ, log_pgsz_fld, \
87 pgsz_shift, page_offset_fld, \
88 scale, page_offset_quantized) \
89 __mlx5_umem_find_best_quantized_pgoff( \
90 umem, \
91 __mlx5_log_page_size_to_bitmap( \
92 __mlx5_bit_sz(typ, log_pgsz_fld), pgsz_shift), \
93 __mlx5_bit_sz(typ, page_offset_fld), \
94 GENMASK(31, order_base_2(scale)), scale, \
95 page_offset_quantized)
96
97#define mlx5_umem_find_best_cq_quantized_pgoff(umem, typ, log_pgsz_fld, \
98 pgsz_shift, page_offset_fld, \
99 scale, page_offset_quantized) \
100 __mlx5_umem_find_best_quantized_pgoff( \
101 umem, \
102 __mlx5_log_page_size_to_bitmap( \
103 __mlx5_bit_sz(typ, log_pgsz_fld), pgsz_shift), \
104 __mlx5_bit_sz(typ, page_offset_fld), 0, scale, \
105 page_offset_quantized)
106
107enum {
108 MLX5_IB_MMAP_OFFSET_START = 9,
109 MLX5_IB_MMAP_OFFSET_END = 255,
110};
111
112enum {
113 MLX5_IB_MMAP_CMD_SHIFT = 8,
114 MLX5_IB_MMAP_CMD_MASK = 0xff,
115};
116
117enum {
118 MLX5_RES_SCAT_DATA32_CQE = 0x1,
119 MLX5_RES_SCAT_DATA64_CQE = 0x2,
120 MLX5_REQ_SCAT_DATA32_CQE = 0x11,
121 MLX5_REQ_SCAT_DATA64_CQE = 0x22,
122};
123
124enum mlx5_ib_mad_ifc_flags {
125 MLX5_MAD_IFC_IGNORE_MKEY = 1,
126 MLX5_MAD_IFC_IGNORE_BKEY = 2,
127 MLX5_MAD_IFC_NET_VIEW = 4,
128};
129
130enum {
131 MLX5_CROSS_CHANNEL_BFREG = 0,
132};
133
134enum {
135 MLX5_CQE_VERSION_V0,
136 MLX5_CQE_VERSION_V1,
137};
138
139enum {
140 MLX5_TM_MAX_RNDV_MSG_SIZE = 64,
141 MLX5_TM_MAX_SGE = 1,
142};
143
144enum {
145 MLX5_IB_INVALID_UAR_INDEX = BIT(31),
146 MLX5_IB_INVALID_BFREG = BIT(31),
147};
148
149enum {
150 MLX5_MAX_MEMIC_PAGES = 0x100,
151 MLX5_MEMIC_ALLOC_SIZE_MASK = 0x3f,
152};
153
154enum {
155 MLX5_MEMIC_BASE_ALIGN = 6,
156 MLX5_MEMIC_BASE_SIZE = 1 << MLX5_MEMIC_BASE_ALIGN,
157};
158
159enum mlx5_ib_mmap_type {
160 MLX5_IB_MMAP_TYPE_MEMIC = 1,
161 MLX5_IB_MMAP_TYPE_VAR = 2,
162 MLX5_IB_MMAP_TYPE_UAR_WC = 3,
163 MLX5_IB_MMAP_TYPE_UAR_NC = 4,
164 MLX5_IB_MMAP_TYPE_MEMIC_OP = 5,
165};
166
167struct mlx5_bfreg_info {
168 u32 *sys_pages;
169 int num_low_latency_bfregs;
170 unsigned int *count;
171
172 /*
173 * protect bfreg allocation data structs
174 */
175 struct mutex lock;
176 u32 ver;
177 u8 lib_uar_4k : 1;
178 u8 lib_uar_dyn : 1;
179 u32 num_sys_pages;
180 u32 num_static_sys_pages;
181 u32 total_num_bfregs;
182 u32 num_dyn_bfregs;
183};
184
185struct mlx5_ib_ucontext {
186 struct ib_ucontext ibucontext;
187 struct list_head db_page_list;
188
189 /* protect doorbell record alloc/free
190 */
191 struct mutex db_page_mutex;
192 struct mlx5_bfreg_info bfregi;
193 u8 cqe_version;
194 /* Transport Domain number */
195 u32 tdn;
196
197 u64 lib_caps;
198 u16 devx_uid;
199 /* For RoCE LAG TX affinity */
200 atomic_t tx_port_affinity;
201};
202
203static inline struct mlx5_ib_ucontext *to_mucontext(struct ib_ucontext *ibucontext)
204{
205 return container_of(ibucontext, struct mlx5_ib_ucontext, ibucontext);
206}
207
208struct mlx5_ib_pd {
209 struct ib_pd ibpd;
210 u32 pdn;
211 u16 uid;
212};
213
214enum {
215 MLX5_IB_FLOW_ACTION_MODIFY_HEADER,
216 MLX5_IB_FLOW_ACTION_PACKET_REFORMAT,
217 MLX5_IB_FLOW_ACTION_DECAP,
218};
219
220#define MLX5_IB_FLOW_MCAST_PRIO (MLX5_BY_PASS_NUM_PRIOS - 1)
221#define MLX5_IB_FLOW_LAST_PRIO (MLX5_BY_PASS_NUM_REGULAR_PRIOS - 1)
222#if (MLX5_IB_FLOW_LAST_PRIO <= 0)
223#error "Invalid number of bypass priorities"
224#endif
225#define MLX5_IB_FLOW_LEFTOVERS_PRIO (MLX5_IB_FLOW_MCAST_PRIO + 1)
226
227#define MLX5_IB_NUM_FLOW_FT (MLX5_IB_FLOW_LEFTOVERS_PRIO + 1)
228#define MLX5_IB_NUM_SNIFFER_FTS 2
229#define MLX5_IB_NUM_EGRESS_FTS 1
230#define MLX5_IB_NUM_FDB_FTS MLX5_BY_PASS_NUM_REGULAR_PRIOS
231
232struct mlx5_ib_anchor {
233 struct mlx5_flow_table *ft;
234 struct mlx5_flow_group *fg_goto_table;
235 struct mlx5_flow_group *fg_drop;
236 struct mlx5_flow_handle *rule_goto_table;
237 struct mlx5_flow_handle *rule_drop;
238 unsigned int rule_goto_table_ref;
239};
240
241struct mlx5_ib_flow_prio {
242 struct mlx5_flow_table *flow_table;
243 struct mlx5_ib_anchor anchor;
244 unsigned int refcount;
245};
246
247struct mlx5_ib_flow_handler {
248 struct list_head list;
249 struct ib_flow ibflow;
250 struct mlx5_ib_flow_prio *prio;
251 struct mlx5_flow_handle *rule;
252 struct ib_counters *ibcounters;
253 struct mlx5_ib_dev *dev;
254 struct mlx5_ib_flow_matcher *flow_matcher;
255};
256
257struct mlx5_ib_flow_matcher {
258 struct mlx5_ib_match_params matcher_mask;
259 int mask_len;
260 enum mlx5_ib_flow_type flow_type;
261 enum mlx5_flow_namespace_type ns_type;
262 u16 priority;
263 struct mlx5_core_dev *mdev;
264 atomic_t usecnt;
265 u8 match_criteria_enable;
266 u32 ib_port;
267};
268
269struct mlx5_ib_steering_anchor {
270 struct mlx5_ib_flow_prio *ft_prio;
271 struct mlx5_ib_dev *dev;
272 atomic_t usecnt;
273};
274
275struct mlx5_ib_pp {
276 u16 index;
277 struct mlx5_core_dev *mdev;
278};
279
280enum mlx5_ib_optional_counter_type {
281 MLX5_IB_OPCOUNTER_CC_RX_CE_PKTS,
282 MLX5_IB_OPCOUNTER_CC_RX_CNP_PKTS,
283 MLX5_IB_OPCOUNTER_CC_TX_CNP_PKTS,
284 MLX5_IB_OPCOUNTER_RDMA_TX_PACKETS,
285 MLX5_IB_OPCOUNTER_RDMA_TX_BYTES,
286 MLX5_IB_OPCOUNTER_RDMA_RX_PACKETS,
287 MLX5_IB_OPCOUNTER_RDMA_RX_BYTES,
288
289 MLX5_IB_OPCOUNTER_CC_RX_CE_PKTS_PER_QP,
290 MLX5_IB_OPCOUNTER_CC_RX_CNP_PKTS_PER_QP,
291 MLX5_IB_OPCOUNTER_CC_TX_CNP_PKTS_PER_QP,
292 MLX5_IB_OPCOUNTER_RDMA_TX_PACKETS_PER_QP,
293 MLX5_IB_OPCOUNTER_RDMA_TX_BYTES_PER_QP,
294 MLX5_IB_OPCOUNTER_RDMA_RX_PACKETS_PER_QP,
295 MLX5_IB_OPCOUNTER_RDMA_RX_BYTES_PER_QP,
296
297 MLX5_IB_OPCOUNTER_MAX,
298};
299
300struct mlx5_ib_flow_db {
301 struct mlx5_ib_flow_prio prios[MLX5_IB_NUM_FLOW_FT];
302 struct mlx5_ib_flow_prio egress_prios[MLX5_IB_NUM_FLOW_FT];
303 struct mlx5_ib_flow_prio sniffer[MLX5_IB_NUM_SNIFFER_FTS];
304 struct mlx5_ib_flow_prio egress[MLX5_IB_NUM_EGRESS_FTS];
305 struct mlx5_ib_flow_prio fdb[MLX5_IB_NUM_FDB_FTS];
306 struct mlx5_ib_flow_prio rdma_rx[MLX5_IB_NUM_FLOW_FT];
307 struct mlx5_ib_flow_prio rdma_tx[MLX5_IB_NUM_FLOW_FT];
308 struct mlx5_ib_flow_prio opfcs[MLX5_IB_OPCOUNTER_MAX];
309 struct mlx5_flow_table *lag_demux_ft;
310 struct mlx5_ib_flow_prio *rdma_transport_rx[MLX5_RDMA_TRANSPORT_BYPASS_PRIO];
311 struct mlx5_ib_flow_prio *rdma_transport_tx[MLX5_RDMA_TRANSPORT_BYPASS_PRIO];
312 /* Protect flow steering bypass flow tables
313 * when add/del flow rules.
314 * only single add/removal of flow steering rule could be done
315 * simultaneously.
316 */
317 struct mutex lock;
318};
319
320/* Use macros here so that don't have to duplicate
321 * enum ib_qp_type for low-level driver
322 */
323
324#define MLX5_IB_QPT_REG_UMR IB_QPT_RESERVED1
325/*
326 * IB_QPT_GSI creates the software wrapper around GSI, and MLX5_IB_QPT_HW_GSI
327 * creates the actual hardware QP.
328 */
329#define MLX5_IB_QPT_HW_GSI IB_QPT_RESERVED2
330#define MLX5_IB_QPT_DCI IB_QPT_RESERVED3
331#define MLX5_IB_QPT_DCT IB_QPT_RESERVED4
332#define MLX5_IB_WR_UMR IB_WR_RESERVED1
333
334#define MLX5_IB_UPD_XLT_ZAP BIT(0)
335#define MLX5_IB_UPD_XLT_ENABLE BIT(1)
336#define MLX5_IB_UPD_XLT_ATOMIC BIT(2)
337#define MLX5_IB_UPD_XLT_ADDR BIT(3)
338#define MLX5_IB_UPD_XLT_PD BIT(4)
339#define MLX5_IB_UPD_XLT_ACCESS BIT(5)
340#define MLX5_IB_UPD_XLT_INDIRECT BIT(6)
341#define MLX5_IB_UPD_XLT_DOWNGRADE BIT(7)
342#define MLX5_IB_UPD_XLT_KEEP_PGSZ BIT(8)
343
344/* Private QP creation flags to be passed in ib_qp_init_attr.create_flags.
345 *
346 * These flags are intended for internal use by the mlx5_ib driver, and they
347 * rely on the range reserved for that use in the ib_qp_create_flags enum.
348 */
349#define MLX5_IB_QP_CREATE_SQPN_QP1 IB_QP_CREATE_RESERVED_START
350
351struct wr_list {
352 u16 opcode;
353 u16 next;
354};
355
356enum mlx5_ib_rq_flags {
357 MLX5_IB_RQ_CVLAN_STRIPPING = 1 << 0,
358 MLX5_IB_RQ_PCI_WRITE_END_PADDING = 1 << 1,
359};
360
361struct mlx5_ib_wq {
362 struct mlx5_frag_buf_ctrl fbc;
363 u64 *wrid;
364 u32 *wr_data;
365 struct wr_list *w_list;
366 unsigned *wqe_head;
367 u16 unsig_count;
368
369 /* serialize post to the work queue
370 */
371 spinlock_t lock;
372 int wqe_cnt;
373 int max_post;
374 int max_gs;
375 int offset;
376 int wqe_shift;
377 unsigned head;
378 unsigned tail;
379 u16 cur_post;
380 u16 last_poll;
381 void *cur_edge;
382};
383
384enum mlx5_ib_wq_flags {
385 MLX5_IB_WQ_FLAGS_DELAY_DROP = 0x1,
386 MLX5_IB_WQ_FLAGS_STRIDING_RQ = 0x2,
387};
388
389#define MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES 9
390#define MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES 16
391#define MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES 6
392#define MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES 13
393#define MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES 3
394
395struct mlx5_ib_rwq {
396 struct ib_wq ibwq;
397 struct mlx5_core_qp core_qp;
398 u32 rq_num_pas;
399 u32 log_rq_stride;
400 u32 log_rq_size;
401 u32 rq_page_offset;
402 u32 log_page_size;
403 u32 log_num_strides;
404 u32 two_byte_shift_en;
405 u32 single_stride_log_num_of_bytes;
406 struct ib_umem *umem;
407 size_t buf_size;
408 unsigned int page_shift;
409 struct mlx5_db db;
410 u32 user_index;
411 u32 wqe_count;
412 u32 wqe_shift;
413 int wq_sig;
414 u32 create_flags; /* Use enum mlx5_ib_wq_flags */
415};
416
417struct mlx5_ib_rwq_ind_table {
418 struct ib_rwq_ind_table ib_rwq_ind_tbl;
419 u32 rqtn;
420 u16 uid;
421};
422
423struct mlx5_ib_ubuffer {
424 struct ib_umem *umem;
425 int buf_size;
426 u64 buf_addr;
427};
428
429struct mlx5_ib_qp_base {
430 struct mlx5_ib_qp *container_mibqp;
431 struct mlx5_core_qp mqp;
432 struct mlx5_ib_ubuffer ubuffer;
433};
434
435struct mlx5_ib_qp_trans {
436 struct mlx5_ib_qp_base base;
437 u16 xrcdn;
438 u32 alt_port;
439 u8 atomic_rd_en;
440 u8 resp_depth;
441};
442
443struct mlx5_ib_rss_qp {
444 u32 tirn;
445};
446
447struct mlx5_ib_rq {
448 struct mlx5_ib_qp_base base;
449 struct mlx5_ib_wq *rq;
450 struct mlx5_ib_ubuffer ubuffer;
451 struct mlx5_db *doorbell;
452 u32 tirn;
453 u8 state;
454 u32 flags;
455};
456
457struct mlx5_ib_sq {
458 struct mlx5_ib_qp_base base;
459 struct mlx5_ib_wq *sq;
460 struct mlx5_ib_ubuffer ubuffer;
461 struct mlx5_db *doorbell;
462 struct mlx5_flow_handle *flow_rule;
463 u32 tisn;
464 u8 state;
465};
466
467struct mlx5_ib_raw_packet_qp {
468 struct mlx5_ib_sq sq;
469 struct mlx5_ib_rq rq;
470};
471
472struct mlx5_bf {
473 int buf_size;
474 unsigned long offset;
475 struct mlx5_sq_bfreg *bfreg;
476};
477
478struct mlx5_ib_dct {
479 struct mlx5_core_dct mdct;
480 u32 *in;
481};
482
483struct mlx5_ib_gsi_qp {
484 struct ib_qp *rx_qp;
485 u32 port_num;
486 struct ib_qp_cap cap;
487 struct ib_cq *cq;
488 struct mlx5_ib_gsi_wr *outstanding_wrs;
489 u32 outstanding_pi, outstanding_ci;
490 int num_qps;
491 /* Protects access to the tx_qps. Post send operations synchronize
492 * with tx_qp creation in setup_qp(). Also protects the
493 * outstanding_wrs array and indices.
494 */
495 spinlock_t lock;
496 struct ib_qp **tx_qps;
497};
498
499struct mlx5_ib_qp {
500 struct ib_qp ibqp;
501 union {
502 struct mlx5_ib_qp_trans trans_qp;
503 struct mlx5_ib_raw_packet_qp raw_packet_qp;
504 struct mlx5_ib_rss_qp rss_qp;
505 struct mlx5_ib_dct dct;
506 struct mlx5_ib_gsi_qp gsi;
507 };
508 struct mlx5_frag_buf buf;
509
510 struct mlx5_db db;
511 struct mlx5_ib_wq rq;
512
513 u8 sq_signal_bits;
514 u8 next_fence;
515 struct mlx5_ib_wq sq;
516
517 /* serialize qp state modifications
518 */
519 struct mutex mutex;
520 /* cached variant of create_flags from struct ib_qp_init_attr */
521 u32 flags;
522 u32 port;
523 u8 state;
524 int max_inline_data;
525 struct mlx5_bf bf;
526 u8 has_rq:1;
527 u8 is_rss:1;
528 u8 is_ooo_rq:1;
529
530 /* only for user space QPs. For kernel
531 * we have it from the bf object
532 */
533 int bfregn;
534
535 struct list_head qps_list;
536 struct list_head cq_recv_list;
537 struct list_head cq_send_list;
538 struct mlx5_rate_limit rl;
539 u32 underlay_qpn;
540 u32 flags_en;
541 /*
542 * IB/core doesn't store low-level QP types, so
543 * store both MLX and IBTA types in the field below.
544 */
545 enum ib_qp_type type;
546 /* A flag to indicate if there's a new counter is configured
547 * but not take effective
548 */
549 u32 counter_pending;
550 u16 gsi_lag_port;
551};
552
553struct mlx5_ib_cq_buf {
554 struct mlx5_frag_buf_ctrl fbc;
555 struct mlx5_frag_buf frag_buf;
556 struct ib_umem *umem;
557 int cqe_size;
558 int nent;
559};
560
561enum mlx5_ib_cq_pr_flags {
562 MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD = 1 << 0,
563 MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS = 1 << 1,
564};
565
566struct mlx5_ib_cq {
567 struct ib_cq ibcq;
568 struct mlx5_core_cq mcq;
569 struct mlx5_ib_cq_buf buf;
570 struct mlx5_db db;
571
572 /* serialize access to the CQ
573 */
574 spinlock_t lock;
575
576 /* protect resize cq
577 */
578 struct mutex resize_mutex;
579 struct mlx5_ib_cq_buf *resize_buf;
580 struct ib_umem *resize_umem;
581 int cqe_size;
582 struct list_head list_send_qp;
583 struct list_head list_recv_qp;
584 u32 create_flags;
585 struct list_head wc_list;
586 enum ib_cq_notify_flags notify_flags;
587 struct work_struct notify_work;
588 u16 private_flags; /* Use mlx5_ib_cq_pr_flags */
589};
590
591struct mlx5_ib_wc {
592 struct ib_wc wc;
593 struct list_head list;
594};
595
596struct mlx5_ib_srq {
597 struct ib_srq ibsrq;
598 struct mlx5_core_srq msrq;
599 struct mlx5_frag_buf buf;
600 struct mlx5_db db;
601 struct mlx5_frag_buf_ctrl fbc;
602 u64 *wrid;
603 /* protect SRQ hanlding
604 */
605 spinlock_t lock;
606 int head;
607 int tail;
608 u16 wqe_ctr;
609 struct ib_umem *umem;
610 /* serialize arming a SRQ
611 */
612 struct mutex mutex;
613 int wq_sig;
614};
615
616struct mlx5_ib_xrcd {
617 struct ib_xrcd ibxrcd;
618 u32 xrcdn;
619};
620
621enum mlx5_ib_mtt_access_flags {
622 MLX5_IB_MTT_READ = (1 << 0),
623 MLX5_IB_MTT_WRITE = (1 << 1),
624};
625
626struct mlx5_user_mmap_entry {
627 struct rdma_user_mmap_entry rdma_entry;
628 u8 mmap_flag;
629 u64 address;
630 u32 page_idx;
631};
632
633enum mlx5_mkey_type {
634 MLX5_MKEY_MR = 1,
635 MLX5_MKEY_MW,
636 MLX5_MKEY_INDIRECT_DEVX,
637 MLX5_MKEY_NULL,
638 MLX5_MKEY_IMPLICIT_CHILD,
639};
640
641/* Used for non-existent ph value */
642#define MLX5_IB_NO_PH 0xff
643
644struct mlx5r_cache_rb_key {
645 u8 ats:1;
646 u8 ph;
647 u16 st_index;
648 unsigned int access_mode;
649 unsigned int access_flags;
650 unsigned int ndescs;
651};
652
653struct mlx5_ib_mkey {
654 u32 key;
655 enum mlx5_mkey_type type;
656 unsigned int ndescs;
657 struct wait_queue_head wait;
658 refcount_t usecount;
659 /* Cacheable user Mkey must hold either a rb_key or a cache_ent. */
660 struct mlx5r_cache_rb_key rb_key;
661 struct mlx5_cache_ent *cache_ent;
662 u8 cacheable : 1;
663};
664
665#define MLX5_IB_MTT_PRESENT (MLX5_IB_MTT_READ | MLX5_IB_MTT_WRITE)
666
667#define MLX5_IB_DM_MEMIC_ALLOWED_ACCESS (IB_ACCESS_LOCAL_WRITE |\
668 IB_ACCESS_REMOTE_WRITE |\
669 IB_ACCESS_REMOTE_READ |\
670 IB_ACCESS_REMOTE_ATOMIC |\
671 IB_ZERO_BASED)
672
673#define MLX5_IB_DM_SW_ICM_ALLOWED_ACCESS (IB_ACCESS_LOCAL_WRITE |\
674 IB_ACCESS_REMOTE_WRITE |\
675 IB_ACCESS_REMOTE_READ |\
676 IB_ZERO_BASED)
677
678#define mlx5_update_odp_stats(mr, counter_name, value) \
679 atomic64_add(value, &((mr)->odp_stats.counter_name))
680
681#define mlx5_update_odp_stats_with_handled(mr, counter_name, value) \
682 do { \
683 mlx5_update_odp_stats(mr, counter_name, value); \
684 atomic64_add(1, &((mr)->odp_stats.counter_name##_handled)); \
685 } while (0)
686
687struct mlx5_ib_mr {
688 struct ib_mr ibmr;
689 struct mlx5_ib_mkey mmkey;
690
691 struct ib_umem *umem;
692 /* The mr is data direct related */
693 u8 data_direct :1;
694
695 union {
696 /* Used only by kernel MRs (umem == NULL) */
697 struct {
698 void *descs;
699 void *descs_alloc;
700 dma_addr_t desc_map;
701 int max_descs;
702 int desc_size;
703 int access_mode;
704
705 /* For Kernel IB_MR_TYPE_INTEGRITY */
706 struct mlx5_core_sig_ctx *sig;
707 struct mlx5_ib_mr *pi_mr;
708 struct mlx5_ib_mr *klm_mr;
709 struct mlx5_ib_mr *mtt_mr;
710 u64 data_iova;
711 u64 pi_iova;
712 int meta_ndescs;
713 int meta_length;
714 int data_length;
715 };
716
717 /* Used only by User MRs (umem != NULL) */
718 struct {
719 unsigned int page_shift;
720 /* Current access_flags */
721 int access_flags;
722
723 /* For User ODP */
724 struct mlx5_ib_mr *parent;
725 struct xarray implicit_children;
726 union {
727 struct work_struct work;
728 } odp_destroy;
729 struct ib_odp_counters odp_stats;
730 bool is_odp_implicit;
731 /* The affilated data direct crossed mr */
732 struct mlx5_ib_mr *dd_crossed_mr;
733 struct list_head dd_node;
734 u8 revoked :1;
735 /* Indicates previous dmabuf page fault occurred */
736 u8 dmabuf_faulted:1;
737 struct mlx5_ib_mkey null_mmkey;
738 };
739 };
740};
741
742static inline bool is_odp_mr(struct mlx5_ib_mr *mr)
743{
744 return IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) && mr->umem &&
745 mr->umem->is_odp;
746}
747
748static inline bool is_dmabuf_mr(struct mlx5_ib_mr *mr)
749{
750 return IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) && mr->umem &&
751 mr->umem->is_dmabuf;
752}
753
754struct mlx5_ib_mw {
755 struct ib_mw ibmw;
756 struct mlx5_ib_mkey mmkey;
757};
758
759struct mlx5_ib_umr_context {
760 struct ib_cqe cqe;
761 enum ib_wc_status status;
762 struct completion done;
763};
764
765enum {
766 MLX5_UMR_STATE_UNINIT,
767 MLX5_UMR_STATE_ACTIVE,
768 MLX5_UMR_STATE_RECOVER,
769 MLX5_UMR_STATE_ERR,
770};
771
772struct umr_common {
773 struct ib_pd *pd;
774 struct ib_cq *cq;
775 struct ib_qp *qp;
776 /* Protects from UMR QP overflow
777 */
778 struct semaphore sem;
779 /* Protects from using UMR while the UMR is not active
780 */
781 struct mutex lock;
782 unsigned int state;
783 /* Protects from repeat UMR QP creation */
784 struct mutex init_lock;
785};
786
787#define NUM_MKEYS_PER_PAGE \
788 ((PAGE_SIZE - sizeof(struct list_head)) / sizeof(u32))
789
790struct mlx5_mkeys_page {
791 u32 mkeys[NUM_MKEYS_PER_PAGE];
792 struct list_head list;
793};
794static_assert(sizeof(struct mlx5_mkeys_page) == PAGE_SIZE);
795
796struct mlx5_mkeys_queue {
797 struct list_head pages_list;
798 u32 num_pages;
799 unsigned long ci;
800 spinlock_t lock; /* sync list ops */
801};
802
803struct mlx5_cache_ent {
804 struct mlx5_mkeys_queue mkeys_queue;
805 u32 pending;
806
807 char name[4];
808
809 struct rb_node node;
810 struct mlx5r_cache_rb_key rb_key;
811
812 u8 is_tmp:1;
813 u8 disabled:1;
814 u8 fill_to_high_water:1;
815 u8 tmp_cleanup_scheduled:1;
816
817 /*
818 * - limit is the low water mark for stored mkeys, 2* limit is the
819 * upper water mark.
820 */
821 u32 in_use;
822 u32 limit;
823
824 /* Statistics */
825 u32 miss;
826
827 struct mlx5_ib_dev *dev;
828 struct delayed_work dwork;
829};
830
831struct mlx5r_async_create_mkey {
832 union {
833 u32 in[MLX5_ST_SZ_BYTES(create_mkey_in)];
834 u32 out[MLX5_ST_SZ_DW(create_mkey_out)];
835 };
836 struct mlx5_async_work cb_work;
837 struct mlx5_cache_ent *ent;
838 u32 mkey;
839};
840
841struct mlx5_mkey_cache {
842 struct workqueue_struct *wq;
843 struct rb_root rb_root;
844 struct mutex rb_lock;
845 struct dentry *fs_root;
846 unsigned long last_add;
847};
848
849struct mlx5_ib_port_resources {
850 struct mlx5_ib_gsi_qp *gsi;
851 struct work_struct pkey_change_work;
852};
853
854struct mlx5_data_direct_resources {
855 u32 pdn;
856 u32 mkey;
857 u32 mkey_ro;
858 u8 mkey_ro_valid :1;
859};
860
861struct mlx5_ib_resources {
862 struct ib_cq *c0;
863 struct mutex cq_lock;
864 u32 xrcdn0;
865 u32 xrcdn1;
866 struct ib_pd *p0;
867 struct ib_srq *s0;
868 struct ib_srq *s1;
869 struct mutex srq_lock;
870 struct mlx5_ib_port_resources ports[2];
871};
872
873#define MAX_OPFC_RULES 2
874
875struct mlx5_ib_op_fc {
876 struct mlx5_fc *fc;
877 struct mlx5_flow_handle *rule[MAX_OPFC_RULES];
878};
879
880struct mlx5_ib_counters {
881 struct rdma_stat_desc *descs;
882 size_t *offsets;
883 u32 num_q_counters;
884 u32 num_cong_counters;
885 u32 num_ext_ppcnt_counters;
886 u32 num_op_counters;
887 u16 set_id;
888 struct mlx5_ib_op_fc opfcs[MLX5_IB_OPCOUNTER_MAX];
889};
890
891int mlx5_ib_fs_add_op_fc(struct mlx5_ib_dev *dev, u32 port_num,
892 struct mlx5_ib_op_fc *opfc,
893 enum mlx5_ib_optional_counter_type type);
894
895void mlx5_ib_fs_remove_op_fc(struct mlx5_ib_dev *dev,
896 struct mlx5_ib_op_fc *opfc,
897 enum mlx5_ib_optional_counter_type type);
898
899int mlx5r_fs_bind_op_fc(struct ib_qp *qp,
900 struct mlx5_fc *fc_arr[MLX5_IB_OPCOUNTER_MAX],
901 struct xarray *qpn_opfc_xa, u32 port);
902
903void mlx5r_fs_unbind_op_fc(struct ib_qp *qp, struct xarray *qpn_opfc_xa);
904
905void mlx5r_fs_destroy_fcs(struct mlx5_ib_dev *dev,
906 struct mlx5_fc *fc_arr[MLX5_IB_OPCOUNTER_MAX]);
907
908struct mlx5_ib_multiport_info;
909
910struct mlx5_ib_multiport {
911 struct mlx5_ib_multiport_info *mpi;
912 /* To be held when accessing the multiport info */
913 spinlock_t mpi_lock;
914};
915
916struct mlx5_roce {
917 /* Protect mlx5_ib_get_netdev from invoking dev_hold() with a NULL
918 * netdev pointer
919 */
920 struct notifier_block nb;
921 struct netdev_net_notifier nn;
922 struct notifier_block mdev_nb;
923 struct net_device *tracking_netdev;
924 atomic_t tx_port_affinity;
925 enum ib_port_state last_port_state;
926 struct mlx5_ib_dev *dev;
927 u32 native_port_num;
928};
929
930struct mlx5_ib_port {
931 struct mlx5_ib_counters cnts;
932 struct mlx5_ib_multiport mp;
933 struct mlx5_ib_dbg_cc_params *dbg_cc_params;
934 struct mlx5_roce roce;
935 struct mlx5_eswitch_rep *rep;
936#ifdef CONFIG_MLX5_MACSEC
937 struct mlx5_reserved_gids *reserved_gids;
938#endif
939};
940
941struct mlx5_ib_dbg_param {
942 int offset;
943 struct mlx5_ib_dev *dev;
944 struct dentry *dentry;
945 u32 port_num;
946};
947
948enum mlx5_ib_dbg_cc_types {
949 MLX5_IB_DBG_CC_RP_CLAMP_TGT_RATE,
950 MLX5_IB_DBG_CC_RP_CLAMP_TGT_RATE_ATI,
951 MLX5_IB_DBG_CC_RP_TIME_RESET,
952 MLX5_IB_DBG_CC_RP_BYTE_RESET,
953 MLX5_IB_DBG_CC_RP_THRESHOLD,
954 MLX5_IB_DBG_CC_RP_AI_RATE,
955 MLX5_IB_DBG_CC_RP_MAX_RATE,
956 MLX5_IB_DBG_CC_RP_HAI_RATE,
957 MLX5_IB_DBG_CC_RP_MIN_DEC_FAC,
958 MLX5_IB_DBG_CC_RP_MIN_RATE,
959 MLX5_IB_DBG_CC_RP_RATE_TO_SET_ON_FIRST_CNP,
960 MLX5_IB_DBG_CC_RP_DCE_TCP_G,
961 MLX5_IB_DBG_CC_RP_DCE_TCP_RTT,
962 MLX5_IB_DBG_CC_RP_RATE_REDUCE_MONITOR_PERIOD,
963 MLX5_IB_DBG_CC_RP_INITIAL_ALPHA_VALUE,
964 MLX5_IB_DBG_CC_RP_GD,
965 MLX5_IB_DBG_CC_NP_MIN_TIME_BETWEEN_CNPS,
966 MLX5_IB_DBG_CC_NP_CNP_DSCP,
967 MLX5_IB_DBG_CC_NP_CNP_PRIO_MODE,
968 MLX5_IB_DBG_CC_NP_CNP_PRIO,
969 MLX5_IB_DBG_CC_GENERAL_RTT_RESP_DSCP_VALID,
970 MLX5_IB_DBG_CC_GENERAL_RTT_RESP_DSCP,
971 MLX5_IB_DBG_CC_MAX,
972};
973
974struct mlx5_ib_dbg_cc_params {
975 struct dentry *root;
976 struct mlx5_ib_dbg_param params[MLX5_IB_DBG_CC_MAX];
977};
978
979enum {
980 MLX5_MAX_DELAY_DROP_TIMEOUT_MS = 100,
981};
982
983struct mlx5_ib_delay_drop {
984 struct mlx5_ib_dev *dev;
985 struct work_struct delay_drop_work;
986 /* serialize setting of delay drop */
987 struct mutex lock;
988 u32 timeout;
989 bool activate;
990 atomic_t events_cnt;
991 atomic_t rqs_cnt;
992 struct dentry *dir_debugfs;
993};
994
995enum mlx5_ib_stages {
996 MLX5_IB_STAGE_INIT,
997 MLX5_IB_STAGE_FS,
998 MLX5_IB_STAGE_CAPS,
999 MLX5_IB_STAGE_NON_DEFAULT_CB,
1000 MLX5_IB_STAGE_ROCE,
1001 MLX5_IB_STAGE_QP,
1002 MLX5_IB_STAGE_SRQ,
1003 MLX5_IB_STAGE_DEVICE_RESOURCES,
1004 MLX5_IB_STAGE_ODP,
1005 MLX5_IB_STAGE_COUNTERS,
1006 MLX5_IB_STAGE_CONG_DEBUGFS,
1007 MLX5_IB_STAGE_BFREG,
1008 MLX5_IB_STAGE_PRE_IB_REG_UMR,
1009 MLX5_IB_STAGE_WHITELIST_UID,
1010 MLX5_IB_STAGE_IB_REG,
1011 MLX5_IB_STAGE_DEVICE_NOTIFIER,
1012 MLX5_IB_STAGE_POST_IB_REG_UMR,
1013 MLX5_IB_STAGE_DELAY_DROP,
1014 MLX5_IB_STAGE_RESTRACK,
1015 MLX5_IB_STAGE_MAX,
1016};
1017
1018struct mlx5_ib_stage {
1019 int (*init)(struct mlx5_ib_dev *dev);
1020 void (*cleanup)(struct mlx5_ib_dev *dev);
1021};
1022
1023#define STAGE_CREATE(_stage, _init, _cleanup) \
1024 .stage[_stage] = {.init = _init, .cleanup = _cleanup}
1025
1026struct mlx5_ib_profile {
1027 struct mlx5_ib_stage stage[MLX5_IB_STAGE_MAX];
1028};
1029
1030struct mlx5_ib_multiport_info {
1031 struct list_head list;
1032 struct mlx5_ib_dev *ibdev;
1033 struct mlx5_core_dev *mdev;
1034 struct notifier_block mdev_events;
1035 struct completion unref_comp;
1036 u64 sys_image_guid;
1037 u32 mdev_refcnt;
1038 bool is_master;
1039 bool unaffiliate;
1040};
1041
1042struct mlx5_ib_flow_action {
1043 struct ib_flow_action ib_action;
1044 union {
1045 struct {
1046 u64 ib_flags;
1047 struct mlx5_accel_esp_xfrm *ctx;
1048 } esp_aes_gcm;
1049 struct {
1050 struct mlx5_ib_dev *dev;
1051 u32 sub_type;
1052 union {
1053 struct mlx5_modify_hdr *modify_hdr;
1054 struct mlx5_pkt_reformat *pkt_reformat;
1055 };
1056 } flow_action_raw;
1057 };
1058};
1059
1060struct mlx5_dm {
1061 struct mlx5_core_dev *dev;
1062 /* This lock is used to protect the access to the shared
1063 * allocation map when concurrent requests by different
1064 * processes are handled.
1065 */
1066 spinlock_t lock;
1067 DECLARE_BITMAP(memic_alloc_pages, MLX5_MAX_MEMIC_PAGES);
1068};
1069
1070struct mlx5_read_counters_attr {
1071 struct mlx5_fc *hw_cntrs_hndl;
1072 u64 *out;
1073 u32 flags;
1074};
1075
1076enum mlx5_ib_counters_type {
1077 MLX5_IB_COUNTERS_FLOW,
1078};
1079
1080struct mlx5_ib_mcounters {
1081 struct ib_counters ibcntrs;
1082 enum mlx5_ib_counters_type type;
1083 /* number of counters supported for this counters type */
1084 u32 counters_num;
1085 struct mlx5_fc *hw_cntrs_hndl;
1086 /* read function for this counters type */
1087 int (*read_counters)(struct ib_device *ibdev,
1088 struct mlx5_read_counters_attr *read_attr);
1089 /* max index set as part of create_flow */
1090 u32 cntrs_max_index;
1091 /* number of counters data entries (<description,index> pair) */
1092 u32 ncounters;
1093 /* counters data array for descriptions and indexes */
1094 struct mlx5_ib_flow_counters_desc *counters_data;
1095 /* protects access to mcounters internal data */
1096 struct mutex mcntrs_mutex;
1097};
1098
1099static inline struct mlx5_ib_mcounters *
1100to_mcounters(struct ib_counters *ibcntrs)
1101{
1102 return container_of(ibcntrs, struct mlx5_ib_mcounters, ibcntrs);
1103}
1104
1105int parse_flow_flow_action(struct mlx5_ib_flow_action *maction,
1106 bool is_egress,
1107 struct mlx5_flow_act *action);
1108struct mlx5_ib_lb_state {
1109 /* protect the user_td */
1110 struct mutex mutex;
1111 u32 user_td;
1112 int qps;
1113 bool enabled;
1114 bool force_enable;
1115};
1116
1117struct mlx5_ib_pf_eq {
1118 struct notifier_block irq_nb;
1119 struct mlx5_ib_dev *dev;
1120 struct mlx5_eq *core;
1121 struct work_struct work;
1122 spinlock_t lock; /* Pagefaults spinlock */
1123 struct workqueue_struct *wq;
1124 mempool_t *pool;
1125};
1126
1127struct mlx5_devx_event_table {
1128 struct mlx5_nb devx_nb;
1129 /* serialize updating the event_xa */
1130 struct mutex event_xa_lock;
1131 struct xarray event_xa;
1132};
1133
1134struct mlx5_var_table {
1135 /* serialize updating the bitmap */
1136 struct mutex bitmap_lock;
1137 unsigned long *bitmap;
1138 u64 hw_start_addr;
1139 u32 stride_size;
1140 u64 num_var_hw_entries;
1141};
1142
1143struct mlx5_port_caps {
1144 bool has_smi;
1145 u8 ext_port_cap;
1146};
1147
1148
1149struct mlx5_special_mkeys {
1150 u32 dump_fill_mkey;
1151 __be32 null_mkey;
1152 __be32 terminate_scatter_list_mkey;
1153};
1154
1155struct mlx5_macsec {
1156 struct mutex lock; /* Protects mlx5_macsec internal contexts */
1157 struct list_head macsec_devices_list;
1158 struct notifier_block blocking_events_nb;
1159};
1160
1161struct mlx5_ib_dev {
1162 struct ib_device ib_dev;
1163 struct mlx5_core_dev *mdev;
1164 struct mlx5_data_direct_dev *data_direct_dev;
1165 /* protect accessing data_direct_dev */
1166 struct mutex data_direct_lock;
1167 struct notifier_block mdev_events;
1168 struct notifier_block lag_events;
1169 int num_ports;
1170 /* serialize update of capability mask
1171 */
1172 struct mutex cap_mask_mutex;
1173 u8 ib_active:1;
1174 u8 is_rep:1;
1175 u8 lag_active:1;
1176 u8 fill_delay;
1177 struct umr_common umrc;
1178 /* sync used page count stats
1179 */
1180 struct mlx5_ib_resources devr;
1181
1182 atomic_t mkey_var;
1183 struct mlx5_mkey_cache cache;
1184 struct timer_list delay_timer;
1185 /* Prevents soft lock on massive reg MRs */
1186 struct mutex slow_path_mutex;
1187 struct ib_odp_caps odp_caps;
1188 u64 odp_max_size;
1189 struct mutex odp_eq_mutex;
1190 struct mlx5_ib_pf_eq odp_pf_eq;
1191
1192 struct xarray odp_mkeys;
1193
1194 struct mlx5_ib_flow_db *flow_db;
1195 /* protect resources needed as part of reset flow */
1196 spinlock_t reset_flow_resource_lock;
1197 struct list_head qp_list;
1198 struct list_head data_direct_mr_list;
1199 /* Array with num_ports elements */
1200 struct mlx5_ib_port *port;
1201 struct mlx5_sq_bfreg bfreg;
1202 struct mlx5_sq_bfreg fp_bfreg;
1203 struct mlx5_ib_delay_drop delay_drop;
1204 const struct mlx5_ib_profile *profile;
1205
1206 struct mlx5_ib_lb_state lb;
1207 u8 umr_fence;
1208 struct list_head ib_dev_list;
1209 u64 sys_image_guid;
1210 struct mlx5_dm dm;
1211 u16 devx_whitelist_uid;
1212 struct mlx5_srq_table srq_table;
1213 struct mlx5_qp_table qp_table;
1214 struct mlx5_async_ctx async_ctx;
1215 struct mlx5_devx_event_table devx_event_table;
1216 struct mlx5_var_table var_table;
1217
1218 struct xarray sig_mrs;
1219 struct mlx5_port_caps port_caps[MLX5_MAX_PORTS];
1220 u16 pkey_table_len;
1221 u8 lag_ports;
1222 struct mlx5_special_mkeys mkeys;
1223 struct mlx5_data_direct_resources ddr;
1224
1225#ifdef CONFIG_MLX5_MACSEC
1226 struct mlx5_macsec macsec;
1227#endif
1228
1229 u8 num_plane;
1230 struct mlx5_ib_dev *smi_dev;
1231 const char *sub_dev_name;
1232};
1233
1234static inline struct mlx5_ib_cq *to_mibcq(struct mlx5_core_cq *mcq)
1235{
1236 return container_of(mcq, struct mlx5_ib_cq, mcq);
1237}
1238
1239static inline struct mlx5_ib_xrcd *to_mxrcd(struct ib_xrcd *ibxrcd)
1240{
1241 return container_of(ibxrcd, struct mlx5_ib_xrcd, ibxrcd);
1242}
1243
1244static inline struct mlx5_ib_dev *to_mdev(struct ib_device *ibdev)
1245{
1246 return container_of(ibdev, struct mlx5_ib_dev, ib_dev);
1247}
1248
1249static inline struct mlx5_ib_dev *mr_to_mdev(struct mlx5_ib_mr *mr)
1250{
1251 return to_mdev(ibdev: mr->ibmr.device);
1252}
1253
1254static inline struct mlx5_ib_dev *mlx5_udata_to_mdev(struct ib_udata *udata)
1255{
1256 struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
1257 udata, struct mlx5_ib_ucontext, ibucontext);
1258
1259 return to_mdev(ibdev: context->ibucontext.device);
1260}
1261
1262static inline struct mlx5_ib_cq *to_mcq(struct ib_cq *ibcq)
1263{
1264 return container_of(ibcq, struct mlx5_ib_cq, ibcq);
1265}
1266
1267static inline struct mlx5_ib_qp *to_mibqp(struct mlx5_core_qp *mqp)
1268{
1269 return container_of(mqp, struct mlx5_ib_qp_base, mqp)->container_mibqp;
1270}
1271
1272static inline struct mlx5_ib_rwq *to_mibrwq(struct mlx5_core_qp *core_qp)
1273{
1274 return container_of(core_qp, struct mlx5_ib_rwq, core_qp);
1275}
1276
1277static inline struct mlx5_ib_pd *to_mpd(struct ib_pd *ibpd)
1278{
1279 return container_of(ibpd, struct mlx5_ib_pd, ibpd);
1280}
1281
1282static inline struct mlx5_ib_srq *to_msrq(struct ib_srq *ibsrq)
1283{
1284 return container_of(ibsrq, struct mlx5_ib_srq, ibsrq);
1285}
1286
1287static inline struct mlx5_ib_qp *to_mqp(struct ib_qp *ibqp)
1288{
1289 return container_of(ibqp, struct mlx5_ib_qp, ibqp);
1290}
1291
1292static inline struct mlx5_ib_rwq *to_mrwq(struct ib_wq *ibwq)
1293{
1294 return container_of(ibwq, struct mlx5_ib_rwq, ibwq);
1295}
1296
1297static inline struct mlx5_ib_rwq_ind_table *to_mrwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
1298{
1299 return container_of(ib_rwq_ind_tbl, struct mlx5_ib_rwq_ind_table, ib_rwq_ind_tbl);
1300}
1301
1302static inline struct mlx5_ib_srq *to_mibsrq(struct mlx5_core_srq *msrq)
1303{
1304 return container_of(msrq, struct mlx5_ib_srq, msrq);
1305}
1306
1307static inline struct mlx5_ib_mr *to_mmr(struct ib_mr *ibmr)
1308{
1309 return container_of(ibmr, struct mlx5_ib_mr, ibmr);
1310}
1311
1312static inline struct mlx5_ib_mw *to_mmw(struct ib_mw *ibmw)
1313{
1314 return container_of(ibmw, struct mlx5_ib_mw, ibmw);
1315}
1316
1317static inline struct mlx5_ib_flow_action *
1318to_mflow_act(struct ib_flow_action *ibact)
1319{
1320 return container_of(ibact, struct mlx5_ib_flow_action, ib_action);
1321}
1322
1323static inline struct mlx5_user_mmap_entry *
1324to_mmmap(struct rdma_user_mmap_entry *rdma_entry)
1325{
1326 return container_of(rdma_entry,
1327 struct mlx5_user_mmap_entry, rdma_entry);
1328}
1329
1330int mlx5_ib_dev_res_cq_init(struct mlx5_ib_dev *dev);
1331int mlx5_ib_dev_res_srq_init(struct mlx5_ib_dev *dev);
1332int mlx5_ib_db_map_user(struct mlx5_ib_ucontext *context, unsigned long virt,
1333 struct mlx5_db *db);
1334void mlx5_ib_db_unmap_user(struct mlx5_ib_ucontext *context, struct mlx5_db *db);
1335void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq);
1336void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq);
1337void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index);
1338int mlx5_ib_create_ah(struct ib_ah *ah, struct rdma_ah_init_attr *init_attr,
1339 struct ib_udata *udata);
1340int mlx5_ib_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr);
1341static inline int mlx5_ib_destroy_ah(struct ib_ah *ah, u32 flags)
1342{
1343 return 0;
1344}
1345int mlx5_ib_create_srq(struct ib_srq *srq, struct ib_srq_init_attr *init_attr,
1346 struct ib_udata *udata);
1347int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
1348 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata);
1349int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr);
1350int mlx5_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata);
1351int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
1352 const struct ib_recv_wr **bad_wr);
1353int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp);
1354void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp);
1355int mlx5_ib_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
1356 struct ib_udata *udata);
1357int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1358 int attr_mask, struct ib_udata *udata);
1359int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
1360 struct ib_qp_init_attr *qp_init_attr);
1361int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata);
1362void mlx5_ib_drain_sq(struct ib_qp *qp);
1363void mlx5_ib_drain_rq(struct ib_qp *qp);
1364int mlx5_ib_read_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
1365 size_t buflen, size_t *bc);
1366int mlx5_ib_read_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
1367 size_t buflen, size_t *bc);
1368int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
1369 size_t buflen, size_t *bc);
1370int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1371 struct uverbs_attr_bundle *attrs);
1372int mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata);
1373int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
1374int mlx5_ib_pre_destroy_cq(struct ib_cq *cq);
1375void mlx5_ib_post_destroy_cq(struct ib_cq *cq);
1376int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags);
1377int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period);
1378int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata);
1379struct ib_mr *mlx5_ib_get_dma_mr(struct ib_pd *pd, int acc);
1380struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1381 u64 virt_addr, int access_flags,
1382 struct ib_dmah *dmah,
1383 struct ib_udata *udata);
1384struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 start,
1385 u64 length, u64 virt_addr,
1386 int fd, int access_flags,
1387 struct ib_dmah *dmah,
1388 struct uverbs_attr_bundle *attrs);
1389int mlx5_ib_advise_mr(struct ib_pd *pd,
1390 enum ib_uverbs_advise_mr_advice advice,
1391 u32 flags,
1392 struct ib_sge *sg_list,
1393 u32 num_sge,
1394 struct uverbs_attr_bundle *attrs);
1395int mlx5_ib_alloc_mw(struct ib_mw *mw, struct ib_udata *udata);
1396int mlx5_ib_dealloc_mw(struct ib_mw *mw);
1397struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
1398 int access_flags);
1399void mlx5_ib_free_odp_mr(struct mlx5_ib_mr *mr);
1400struct ib_mr *mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
1401 u64 length, u64 virt_addr, int access_flags,
1402 struct ib_pd *pd, struct ib_udata *udata);
1403int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata);
1404struct ib_mr *mlx5_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1405 u32 max_num_sg);
1406struct ib_mr *mlx5_ib_alloc_mr_integrity(struct ib_pd *pd,
1407 u32 max_num_sg,
1408 u32 max_num_meta_sg);
1409int mlx5_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
1410 unsigned int *sg_offset);
1411int mlx5_ib_map_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
1412 int data_sg_nents, unsigned int *data_sg_offset,
1413 struct scatterlist *meta_sg, int meta_sg_nents,
1414 unsigned int *meta_sg_offset);
1415int mlx5_ib_process_mad(struct ib_device *ibdev, int mad_flags, u32 port_num,
1416 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
1417 const struct ib_mad *in, struct ib_mad *out,
1418 size_t *out_mad_size, u16 *out_mad_pkey_index);
1419int mlx5_ib_alloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata);
1420int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata);
1421int mlx5_query_ext_port_caps(struct mlx5_ib_dev *dev, unsigned int port);
1422int mlx5_query_mad_ifc_system_image_guid(struct ib_device *ibdev,
1423 __be64 *sys_image_guid);
1424int mlx5_query_mad_ifc_max_pkeys(struct ib_device *ibdev,
1425 u16 *max_pkeys);
1426int mlx5_query_mad_ifc_vendor_id(struct ib_device *ibdev,
1427 u32 *vendor_id);
1428int mlx5_query_mad_ifc_node_desc(struct mlx5_ib_dev *dev, char *node_desc);
1429int mlx5_query_mad_ifc_node_guid(struct mlx5_ib_dev *dev, __be64 *node_guid);
1430int mlx5_query_mad_ifc_pkey(struct ib_device *ibdev, u32 port, u16 index,
1431 u16 *pkey);
1432int mlx5_query_mad_ifc_gids(struct ib_device *ibdev, u32 port, int index,
1433 union ib_gid *gid);
1434int mlx5_query_mad_ifc_port(struct ib_device *ibdev, u32 port,
1435 struct ib_port_attr *props);
1436int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1437 struct ib_port_attr *props);
1438void mlx5_ib_populate_pas(struct ib_umem *umem, size_t page_size, __be64 *pas,
1439 u64 access_flags);
1440int mlx5_ib_get_cqe_size(struct ib_cq *ibcq);
1441int mlx5_mkey_cache_init(struct mlx5_ib_dev *dev);
1442void mlx5_mkey_cache_cleanup(struct mlx5_ib_dev *dev);
1443struct mlx5_cache_ent *
1444mlx5r_cache_create_ent_locked(struct mlx5_ib_dev *dev,
1445 struct mlx5r_cache_rb_key rb_key,
1446 bool persistent_entry);
1447
1448struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
1449 int access_flags, int access_mode,
1450 int ndescs);
1451
1452int mlx5_ib_check_mr_status(struct ib_mr *ibmr, u32 check_mask,
1453 struct ib_mr_status *mr_status);
1454struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
1455 struct ib_wq_init_attr *init_attr,
1456 struct ib_udata *udata);
1457int mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata);
1458int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1459 u32 wq_attr_mask, struct ib_udata *udata);
1460int mlx5_ib_create_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_table,
1461 struct ib_rwq_ind_table_init_attr *init_attr,
1462 struct ib_udata *udata);
1463int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *wq_ind_table);
1464struct ib_mr *mlx5_ib_reg_dm_mr(struct ib_pd *pd, struct ib_dm *dm,
1465 struct ib_dm_mr_attr *attr,
1466 struct uverbs_attr_bundle *attrs);
1467void mlx5_ib_data_direct_bind(struct mlx5_ib_dev *ibdev,
1468 struct mlx5_data_direct_dev *dev);
1469void mlx5_ib_data_direct_unbind(struct mlx5_ib_dev *ibdev);
1470void mlx5_ib_revoke_data_direct_mrs(struct mlx5_ib_dev *dev);
1471
1472#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1473int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev);
1474int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq);
1475void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *ibdev);
1476int __init mlx5_ib_odp_init(void);
1477void mlx5_ib_odp_cleanup(void);
1478int mlx5_odp_init_mkey_cache(struct mlx5_ib_dev *dev);
1479int mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
1480 struct mlx5_ib_mr *mr, int flags);
1481
1482int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1483 enum ib_uverbs_advise_mr_advice advice,
1484 u32 flags, struct ib_sge *sg_list, u32 num_sge);
1485int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr);
1486int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr);
1487#else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
1488static inline int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev) { return 0; }
1489static inline int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev,
1490 struct mlx5_ib_pf_eq *eq)
1491{
1492 return 0;
1493}
1494static inline void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *ibdev) {}
1495static inline int mlx5_ib_odp_init(void) { return 0; }
1496static inline void mlx5_ib_odp_cleanup(void) {}
1497static inline int mlx5_odp_init_mkey_cache(struct mlx5_ib_dev *dev)
1498{
1499 return 0;
1500}
1501static inline int mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
1502 struct mlx5_ib_mr *mr, int flags)
1503{
1504 return -EOPNOTSUPP;
1505}
1506
1507static inline int
1508mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
1509 enum ib_uverbs_advise_mr_advice advice, u32 flags,
1510 struct ib_sge *sg_list, u32 num_sge)
1511{
1512 return -EOPNOTSUPP;
1513}
1514static inline int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr)
1515{
1516 return -EOPNOTSUPP;
1517}
1518static inline int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr)
1519{
1520 return -EOPNOTSUPP;
1521}
1522#endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
1523
1524extern const struct mmu_interval_notifier_ops mlx5_mn_ops;
1525
1526/* Needed for rep profile */
1527void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
1528 const struct mlx5_ib_profile *profile,
1529 int stage);
1530int __mlx5_ib_add(struct mlx5_ib_dev *dev,
1531 const struct mlx5_ib_profile *profile);
1532
1533int mlx5_ib_get_vf_config(struct ib_device *device, int vf,
1534 u32 port, struct ifla_vf_info *info);
1535int mlx5_ib_set_vf_link_state(struct ib_device *device, int vf,
1536 u32 port, int state);
1537int mlx5_ib_get_vf_stats(struct ib_device *device, int vf,
1538 u32 port, struct ifla_vf_stats *stats);
1539int mlx5_ib_get_vf_guid(struct ib_device *device, int vf, u32 port,
1540 struct ifla_vf_guid *node_guid,
1541 struct ifla_vf_guid *port_guid);
1542int mlx5_ib_set_vf_guid(struct ib_device *device, int vf, u32 port,
1543 u64 guid, int type);
1544
1545__be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
1546 const struct ib_gid_attr *attr);
1547
1548void mlx5_ib_cleanup_cong_debugfs(struct mlx5_ib_dev *dev, u32 port_num);
1549void mlx5_ib_init_cong_debugfs(struct mlx5_ib_dev *dev, u32 port_num);
1550
1551/* GSI QP helper functions */
1552int mlx5_ib_create_gsi(struct ib_pd *pd, struct mlx5_ib_qp *mqp,
1553 struct ib_qp_init_attr *attr);
1554int mlx5_ib_destroy_gsi(struct mlx5_ib_qp *mqp);
1555int mlx5_ib_gsi_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1556 int attr_mask);
1557int mlx5_ib_gsi_query_qp(struct ib_qp *qp, struct ib_qp_attr *qp_attr,
1558 int qp_attr_mask,
1559 struct ib_qp_init_attr *qp_init_attr);
1560int mlx5_ib_gsi_post_send(struct ib_qp *qp, const struct ib_send_wr *wr,
1561 const struct ib_send_wr **bad_wr);
1562int mlx5_ib_gsi_post_recv(struct ib_qp *qp, const struct ib_recv_wr *wr,
1563 const struct ib_recv_wr **bad_wr);
1564void mlx5_ib_gsi_pkey_change(struct mlx5_ib_gsi_qp *gsi);
1565
1566int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc);
1567
1568void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi,
1569 int bfregn);
1570struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi);
1571struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *dev,
1572 u32 ib_port_num,
1573 u32 *native_port_num);
1574void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *dev,
1575 u32 port_num);
1576
1577extern const struct uapi_definition mlx5_ib_devx_defs[];
1578extern const struct uapi_definition mlx5_ib_flow_defs[];
1579extern const struct uapi_definition mlx5_ib_qos_defs[];
1580extern const struct uapi_definition mlx5_ib_std_types_defs[];
1581extern const struct uapi_definition mlx5_ib_create_cq_defs[];
1582
1583static inline int is_qp1(enum ib_qp_type qp_type)
1584{
1585 return qp_type == MLX5_IB_QPT_HW_GSI || qp_type == IB_QPT_GSI;
1586}
1587
1588static inline u32 check_cq_create_flags(u32 flags)
1589{
1590 /*
1591 * It returns non-zero value for unsupported CQ
1592 * create flags, otherwise it returns zero.
1593 */
1594 return (flags & ~(IB_UVERBS_CQ_FLAGS_IGNORE_OVERRUN |
1595 IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION));
1596}
1597
1598static inline int verify_assign_uidx(u8 cqe_version, u32 cmd_uidx,
1599 u32 *user_index)
1600{
1601 if (cqe_version) {
1602 if ((cmd_uidx == MLX5_IB_DEFAULT_UIDX) ||
1603 (cmd_uidx & ~MLX5_USER_ASSIGNED_UIDX_MASK))
1604 return -EINVAL;
1605 *user_index = cmd_uidx;
1606 } else {
1607 *user_index = MLX5_IB_DEFAULT_UIDX;
1608 }
1609
1610 return 0;
1611}
1612
1613static inline int get_qp_user_index(struct mlx5_ib_ucontext *ucontext,
1614 struct mlx5_ib_create_qp *ucmd,
1615 int inlen,
1616 u32 *user_index)
1617{
1618 u8 cqe_version = ucontext->cqe_version;
1619
1620 if ((offsetofend(typeof(*ucmd), uidx) <= inlen) && !cqe_version &&
1621 (ucmd->uidx == MLX5_IB_DEFAULT_UIDX))
1622 return 0;
1623
1624 if ((offsetofend(typeof(*ucmd), uidx) <= inlen) != !!cqe_version)
1625 return -EINVAL;
1626
1627 return verify_assign_uidx(cqe_version, cmd_uidx: ucmd->uidx, user_index);
1628}
1629
1630static inline int get_srq_user_index(struct mlx5_ib_ucontext *ucontext,
1631 struct mlx5_ib_create_srq *ucmd,
1632 int inlen,
1633 u32 *user_index)
1634{
1635 u8 cqe_version = ucontext->cqe_version;
1636
1637 if ((offsetofend(typeof(*ucmd), uidx) <= inlen) && !cqe_version &&
1638 (ucmd->uidx == MLX5_IB_DEFAULT_UIDX))
1639 return 0;
1640
1641 if ((offsetofend(typeof(*ucmd), uidx) <= inlen) != !!cqe_version)
1642 return -EINVAL;
1643
1644 return verify_assign_uidx(cqe_version, cmd_uidx: ucmd->uidx, user_index);
1645}
1646
1647static inline int get_uars_per_sys_page(struct mlx5_ib_dev *dev, bool lib_support)
1648{
1649 return lib_support && MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1650 MLX5_UARS_IN_PAGE : 1;
1651}
1652
1653extern void *xlt_emergency_page;
1654
1655int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
1656 struct mlx5_bfreg_info *bfregi, u32 bfregn,
1657 bool dyn_bfreg);
1658
1659static inline int mlx5r_store_odp_mkey(struct mlx5_ib_dev *dev,
1660 struct mlx5_ib_mkey *mmkey)
1661{
1662 refcount_set(r: &mmkey->usecount, n: 1);
1663
1664 return xa_err(entry: xa_store(&dev->odp_mkeys, index: mlx5_base_mkey(key: mmkey->key),
1665 entry: mmkey, GFP_KERNEL));
1666}
1667
1668/* deref an mkey that can participate in ODP flow */
1669static inline void mlx5r_deref_odp_mkey(struct mlx5_ib_mkey *mmkey)
1670{
1671 if (refcount_dec_and_test(r: &mmkey->usecount))
1672 wake_up(&mmkey->wait);
1673}
1674
1675/* deref an mkey that can participate in ODP flow and wait for relese */
1676static inline void mlx5r_deref_wait_odp_mkey(struct mlx5_ib_mkey *mmkey)
1677{
1678 mlx5r_deref_odp_mkey(mmkey);
1679 wait_event(mmkey->wait, refcount_read(&mmkey->usecount) == 0);
1680}
1681
1682static inline bool mlx5_ib_lag_should_assign_affinity(struct mlx5_ib_dev *dev)
1683{
1684 /*
1685 * If the driver is in hash mode and the port_select_flow_table_bypass cap
1686 * is supported, it means that the driver no longer needs to assign the port
1687 * affinity by default. If a user wants to set the port affinity explicitly,
1688 * the user has a dedicated API to do that, so there is no need to assign
1689 * the port affinity by default.
1690 */
1691 if (dev->lag_active &&
1692 mlx5_lag_mode_is_hash(dev: dev->mdev) &&
1693 MLX5_CAP_PORT_SELECTION(dev->mdev, port_select_flow_table_bypass))
1694 return 0;
1695
1696 if (mlx5_lag_is_lacp_owner(dev: dev->mdev) && !dev->lag_active)
1697 return 0;
1698
1699 return dev->lag_active ||
1700 (MLX5_CAP_GEN(dev->mdev, num_lag_ports) > 1 &&
1701 MLX5_CAP_GEN(dev->mdev, lag_tx_port_affinity));
1702}
1703
1704static inline bool rt_supported(int ts_cap)
1705{
1706 return ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_REAL_TIME ||
1707 ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
1708}
1709
1710/*
1711 * PCI Peer to Peer is a trainwreck. If no switch is present then things
1712 * sometimes work, depending on the pci_distance_p2p logic for excluding broken
1713 * root complexes. However if a switch is present in the path, then things get
1714 * really ugly depending on how the switch is setup. This table assumes that the
1715 * root complex is strict and is validating that all req/reps are matches
1716 * perfectly - so any scenario where it sees only half the transaction is a
1717 * failure.
1718 *
1719 * CR/RR/DT ATS RO P2P
1720 * 00X X X OK
1721 * 010 X X fails (request is routed to root but root never sees comp)
1722 * 011 0 X fails (request is routed to root but root never sees comp)
1723 * 011 1 X OK
1724 * 10X X 1 OK
1725 * 101 X 0 fails (completion is routed to root but root didn't see req)
1726 * 110 X 0 SLOW
1727 * 111 0 0 SLOW
1728 * 111 1 0 fails (completion is routed to root but root didn't see req)
1729 * 111 1 1 OK
1730 *
1731 * Unfortunately we cannot reliably know if a switch is present or what the
1732 * CR/RR/DT ACS settings are, as in a VM that is all hidden. Assume that
1733 * CR/RR/DT is 111 if the ATS cap is enabled and follow the last three rows.
1734 *
1735 * For now assume if the umem is a dma_buf then it is P2P.
1736 */
1737static inline bool mlx5_umem_needs_ats(struct mlx5_ib_dev *dev,
1738 struct ib_umem *umem, int access_flags)
1739{
1740 if (!MLX5_CAP_GEN(dev->mdev, ats) || !umem->is_dmabuf)
1741 return false;
1742 return access_flags & IB_ACCESS_RELAXED_ORDERING;
1743}
1744
1745int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
1746 unsigned int index, const union ib_gid *gid,
1747 const struct ib_gid_attr *attr);
1748
1749static inline u32 smi_to_native_portnum(struct mlx5_ib_dev *dev, u32 port)
1750{
1751 return (port - 1) / dev->num_ports + 1;
1752}
1753
1754static inline unsigned int get_max_log_entity_size_cap(struct mlx5_ib_dev *dev,
1755 int access_mode)
1756{
1757 int max_log_size = 0;
1758
1759 if (access_mode == MLX5_MKC_ACCESS_MODE_MTT)
1760 max_log_size =
1761 MLX5_CAP_GEN_2(dev->mdev, max_mkey_log_entity_size_mtt);
1762 else if (access_mode == MLX5_MKC_ACCESS_MODE_KSM)
1763 max_log_size = MLX5_CAP_GEN_2(
1764 dev->mdev, max_mkey_log_entity_size_fixed_buffer);
1765
1766 if (!max_log_size ||
1767 (max_log_size > 31 &&
1768 !MLX5_CAP_GEN_2(dev->mdev, umr_log_entity_size_5)))
1769 max_log_size = 31;
1770
1771 return max_log_size;
1772}
1773
1774static inline unsigned int get_min_log_entity_size_cap(struct mlx5_ib_dev *dev,
1775 int access_mode)
1776{
1777 int min_log_size = 0;
1778
1779 if (access_mode == MLX5_MKC_ACCESS_MODE_KSM &&
1780 MLX5_CAP_GEN_2(dev->mdev,
1781 min_mkey_log_entity_size_fixed_buffer_valid))
1782 min_log_size = MLX5_CAP_GEN_2(
1783 dev->mdev, min_mkey_log_entity_size_fixed_buffer);
1784 else
1785 min_log_size =
1786 MLX5_CAP_GEN_2(dev->mdev, log_min_mkey_entity_size);
1787
1788 min_log_size = max(min_log_size, MLX5_ADAPTER_PAGE_SHIFT);
1789 return min_log_size;
1790}
1791
1792/*
1793 * For mkc users, instead of a page_offset the command has a start_iova which
1794 * specifies both the page_offset and the on-the-wire IOVA
1795 */
1796static __always_inline unsigned long
1797mlx5_umem_mkc_find_best_pgsz(struct mlx5_ib_dev *dev, struct ib_umem *umem,
1798 u64 iova, int access_mode)
1799{
1800 unsigned int max_log_entity_size_cap, min_log_entity_size_cap;
1801 unsigned long bitmap;
1802
1803 max_log_entity_size_cap = get_max_log_entity_size_cap(dev, access_mode);
1804 min_log_entity_size_cap = get_min_log_entity_size_cap(dev, access_mode);
1805
1806 bitmap = GENMASK_ULL(max_log_entity_size_cap, min_log_entity_size_cap);
1807
1808 /* In KSM mode HW requires IOVA and mkey's page size to be aligned */
1809 if (access_mode == MLX5_MKC_ACCESS_MODE_KSM && iova)
1810 bitmap &= GENMASK_ULL(__ffs64(iova), 0);
1811
1812 return ib_umem_find_best_pgsz(umem, pgsz_bitmap: bitmap, virt: iova);
1813}
1814
1815static inline unsigned long
1816mlx5_umem_dmabuf_find_best_pgsz(struct ib_umem_dmabuf *umem_dmabuf,
1817 int access_mode)
1818{
1819 return mlx5_umem_mkc_find_best_pgsz(dev: to_mdev(ibdev: umem_dmabuf->umem.ibdev),
1820 umem: &umem_dmabuf->umem,
1821 iova: umem_dmabuf->umem.iova,
1822 access_mode);
1823}
1824
1825#endif /* MLX5_IB_H */
1826

source code of linux/drivers/infiniband/hw/mlx5/mlx5_ib.h