xref: /linux/drivers/infiniband/sw/rxe/rxe_verbs.h (revision 3d807a3e)
1 /* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
2 /*
3  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5  */
6 
7 #ifndef RXE_VERBS_H
8 #define RXE_VERBS_H
9 
10 #include <linux/interrupt.h>
11 #include <linux/workqueue.h>
12 #include "rxe_pool.h"
13 #include "rxe_task.h"
14 #include "rxe_hw_counters.h"
15 
pkey_match(u16 key1,u16 key2)16 static inline int pkey_match(u16 key1, u16 key2)
17 {
18 	return (((key1 & 0x7fff) != 0) &&
19 		((key1 & 0x7fff) == (key2 & 0x7fff)) &&
20 		((key1 & 0x8000) || (key2 & 0x8000))) ? 1 : 0;
21 }
22 
23 /* Return >0 if psn_a > psn_b
24  *	   0 if psn_a == psn_b
25  *	  <0 if psn_a < psn_b
26  */
psn_compare(u32 psn_a,u32 psn_b)27 static inline int psn_compare(u32 psn_a, u32 psn_b)
28 {
29 	s32 diff;
30 
31 	diff = (psn_a - psn_b) << 8;
32 	return diff;
33 }
34 
35 struct rxe_ucontext {
36 	struct ib_ucontext ibuc;
37 	struct rxe_pool_elem	elem;
38 };
39 
40 struct rxe_pd {
41 	struct ib_pd            ibpd;
42 	struct rxe_pool_elem	elem;
43 };
44 
45 struct rxe_ah {
46 	struct ib_ah		ibah;
47 	struct rxe_pool_elem	elem;
48 	struct rxe_av		av;
49 	bool			is_user;
50 	int			ah_num;
51 };
52 
53 struct rxe_cqe {
54 	union {
55 		struct ib_wc		ibwc;
56 		struct ib_uverbs_wc	uibwc;
57 	};
58 };
59 
60 struct rxe_cq {
61 	struct ib_cq		ibcq;
62 	struct rxe_pool_elem	elem;
63 	struct rxe_queue	*queue;
64 	spinlock_t		cq_lock;
65 	u8			notify;
66 	bool			is_user;
67 	atomic_t		num_wq;
68 };
69 
70 enum wqe_state {
71 	wqe_state_posted,
72 	wqe_state_processing,
73 	wqe_state_pending,
74 	wqe_state_done,
75 	wqe_state_error,
76 };
77 
78 struct rxe_sq {
79 	int			max_wr;
80 	int			max_sge;
81 	int			max_inline;
82 	spinlock_t		sq_lock; /* guard queue */
83 	struct rxe_queue	*queue;
84 };
85 
86 struct rxe_rq {
87 	int			max_wr;
88 	int			max_sge;
89 	spinlock_t		producer_lock; /* guard queue producer */
90 	spinlock_t		consumer_lock; /* guard queue consumer */
91 	struct rxe_queue	*queue;
92 };
93 
94 struct rxe_srq {
95 	struct ib_srq		ibsrq;
96 	struct rxe_pool_elem	elem;
97 	struct rxe_pd		*pd;
98 	struct rxe_rq		rq;
99 	u32			srq_num;
100 
101 	int			limit;
102 	int			error;
103 };
104 
105 struct rxe_req_info {
106 	int			wqe_index;
107 	u32			psn;
108 	int			opcode;
109 	atomic_t		rd_atomic;
110 	int			wait_fence;
111 	int			need_rd_atomic;
112 	int			wait_psn;
113 	int			need_retry;
114 	int			wait_for_rnr_timer;
115 	int			noack_pkts;
116 	int			again;
117 };
118 
119 struct rxe_comp_info {
120 	u32			psn;
121 	int			opcode;
122 	int			timeout;
123 	int			timeout_retry;
124 	int			started_retry;
125 	u32			retry_cnt;
126 	u32			rnr_retry;
127 };
128 
129 enum rdatm_res_state {
130 	rdatm_res_state_next,
131 	rdatm_res_state_new,
132 	rdatm_res_state_replay,
133 };
134 
135 struct resp_res {
136 	int			type;
137 	int			replay;
138 	u32			first_psn;
139 	u32			last_psn;
140 	u32			cur_psn;
141 	enum rdatm_res_state	state;
142 
143 	union {
144 		struct {
145 			u64		orig_val;
146 		} atomic;
147 		struct {
148 			u64		va_org;
149 			u32		rkey;
150 			u32		length;
151 			u64		va;
152 			u32		resid;
153 		} read;
154 		struct {
155 			u32		length;
156 			u64		va;
157 			u8		type;
158 			u8		level;
159 		} flush;
160 	};
161 };
162 
163 struct rxe_resp_info {
164 	u32			msn;
165 	u32			psn;
166 	u32			ack_psn;
167 	int			opcode;
168 	int			drop_msg;
169 	int			goto_error;
170 	int			sent_psn_nak;
171 	enum ib_wc_status	status;
172 	u8			aeth_syndrome;
173 
174 	/* Receive only */
175 	struct rxe_recv_wqe	*wqe;
176 
177 	/* RDMA read / atomic only */
178 	u64			va;
179 	u64			offset;
180 	struct rxe_mr		*mr;
181 	u32			resid;
182 	u32			rkey;
183 	u32			length;
184 
185 	/* SRQ only */
186 	struct {
187 		struct rxe_recv_wqe	wqe;
188 		struct ib_sge		sge[RXE_MAX_SGE];
189 	} srq_wqe;
190 
191 	/* Responder resources. It's a circular list where the oldest
192 	 * resource is dropped first.
193 	 */
194 	struct resp_res		*resources;
195 	unsigned int		res_head;
196 	unsigned int		res_tail;
197 	struct resp_res		*res;
198 };
199 
200 struct rxe_qp {
201 	struct ib_qp		ibqp;
202 	struct rxe_pool_elem	elem;
203 	struct ib_qp_attr	attr;
204 	unsigned int		valid;
205 	unsigned int		mtu;
206 	bool			is_user;
207 
208 	struct rxe_pd		*pd;
209 	struct rxe_srq		*srq;
210 	struct rxe_cq		*scq;
211 	struct rxe_cq		*rcq;
212 
213 	enum ib_sig_type	sq_sig_type;
214 
215 	struct rxe_sq		sq;
216 	struct rxe_rq		rq;
217 
218 	struct socket		*sk;
219 	u32			dst_cookie;
220 	u16			src_port;
221 
222 	struct rxe_av		pri_av;
223 	struct rxe_av		alt_av;
224 
225 	atomic_t		mcg_num;
226 
227 	struct sk_buff_head	req_pkts;
228 	struct sk_buff_head	resp_pkts;
229 
230 	struct rxe_task		send_task;
231 	struct rxe_task		recv_task;
232 
233 	struct rxe_req_info	req;
234 	struct rxe_comp_info	comp;
235 	struct rxe_resp_info	resp;
236 
237 	atomic_t		ssn;
238 	atomic_t		skb_out;
239 	int			need_req_skb;
240 
241 	/* Timer for retranmitting packet when ACKs have been lost. RC
242 	 * only. The requester sets it when it is not already
243 	 * started. The responder resets it whenever an ack is
244 	 * received.
245 	 */
246 	struct timer_list retrans_timer;
247 	u64 qp_timeout_jiffies;
248 
249 	/* Timer for handling RNR NAKS. */
250 	struct timer_list rnr_nak_timer;
251 
252 	spinlock_t		state_lock; /* guard requester and completer */
253 
254 	struct execute_work	cleanup_work;
255 };
256 
257 enum {
258 	RXE_ACCESS_REMOTE	= IB_ACCESS_REMOTE_READ
259 				| IB_ACCESS_REMOTE_WRITE
260 				| IB_ACCESS_REMOTE_ATOMIC,
261 	RXE_ACCESS_SUPPORTED_MR	= RXE_ACCESS_REMOTE
262 				| IB_ACCESS_LOCAL_WRITE
263 				| IB_ACCESS_MW_BIND
264 				| IB_ACCESS_ON_DEMAND
265 				| IB_ACCESS_FLUSH_GLOBAL
266 				| IB_ACCESS_FLUSH_PERSISTENT
267 				| IB_ACCESS_OPTIONAL,
268 	RXE_ACCESS_SUPPORTED_QP	= RXE_ACCESS_SUPPORTED_MR,
269 	RXE_ACCESS_SUPPORTED_MW	= RXE_ACCESS_SUPPORTED_MR
270 				| IB_ZERO_BASED,
271 };
272 
273 enum rxe_mr_state {
274 	RXE_MR_STATE_INVALID,
275 	RXE_MR_STATE_FREE,
276 	RXE_MR_STATE_VALID,
277 };
278 
279 enum rxe_mr_copy_dir {
280 	RXE_TO_MR_OBJ,
281 	RXE_FROM_MR_OBJ,
282 };
283 
284 enum rxe_mr_lookup_type {
285 	RXE_LOOKUP_LOCAL,
286 	RXE_LOOKUP_REMOTE,
287 };
288 
289 enum rxe_rereg {
290 	RXE_MR_REREG_SUPPORTED	= IB_MR_REREG_PD
291 				| IB_MR_REREG_ACCESS,
292 };
293 
rkey_is_mw(u32 rkey)294 static inline int rkey_is_mw(u32 rkey)
295 {
296 	u32 index = rkey >> 8;
297 
298 	return (index >= RXE_MIN_MW_INDEX) && (index <= RXE_MAX_MW_INDEX);
299 }
300 
301 struct rxe_mr {
302 	struct rxe_pool_elem	elem;
303 	struct ib_mr		ibmr;
304 
305 	struct ib_umem		*umem;
306 
307 	u32			lkey;
308 	u32			rkey;
309 	enum rxe_mr_state	state;
310 	int			access;
311 	atomic_t		num_mw;
312 
313 	unsigned int		page_offset;
314 	unsigned int		page_shift;
315 	u64			page_mask;
316 
317 	u32			num_buf;
318 	u32			nbuf;
319 
320 	struct xarray		page_list;
321 };
322 
mr_page_size(struct rxe_mr * mr)323 static inline unsigned int mr_page_size(struct rxe_mr *mr)
324 {
325 	return mr ? mr->ibmr.page_size : PAGE_SIZE;
326 }
327 
328 enum rxe_mw_state {
329 	RXE_MW_STATE_INVALID	= RXE_MR_STATE_INVALID,
330 	RXE_MW_STATE_FREE	= RXE_MR_STATE_FREE,
331 	RXE_MW_STATE_VALID	= RXE_MR_STATE_VALID,
332 };
333 
334 struct rxe_mw {
335 	struct ib_mw		ibmw;
336 	struct rxe_pool_elem	elem;
337 	spinlock_t		lock;
338 	enum rxe_mw_state	state;
339 	struct rxe_qp		*qp; /* Type 2 only */
340 	struct rxe_mr		*mr;
341 	u32			rkey;
342 	int			access;
343 	u64			addr;
344 	u64			length;
345 };
346 
347 struct rxe_mcg {
348 	struct rb_node		node;
349 	struct kref		ref_cnt;
350 	struct rxe_dev		*rxe;
351 	struct list_head	qp_list;
352 	union ib_gid		mgid;
353 	atomic_t		qp_num;
354 	u32			qkey;
355 	u16			pkey;
356 };
357 
358 struct rxe_mca {
359 	struct list_head	qp_list;
360 	struct rxe_qp		*qp;
361 };
362 
363 struct rxe_port {
364 	struct ib_port_attr	attr;
365 	__be64			port_guid;
366 	__be64			subnet_prefix;
367 	spinlock_t		port_lock; /* guard port */
368 	unsigned int		mtu_cap;
369 	/* special QPs */
370 	u32			qp_gsi_index;
371 };
372 
373 struct rxe_dev {
374 	struct ib_device	ib_dev;
375 	struct ib_device_attr	attr;
376 	int			max_ucontext;
377 	int			max_inline_data;
378 	struct mutex	usdev_lock;
379 
380 	struct net_device	*ndev;
381 
382 	struct rxe_pool		uc_pool;
383 	struct rxe_pool		pd_pool;
384 	struct rxe_pool		ah_pool;
385 	struct rxe_pool		srq_pool;
386 	struct rxe_pool		qp_pool;
387 	struct rxe_pool		cq_pool;
388 	struct rxe_pool		mr_pool;
389 	struct rxe_pool		mw_pool;
390 
391 	/* multicast support */
392 	spinlock_t		mcg_lock;
393 	struct rb_root		mcg_tree;
394 	atomic_t		mcg_num;
395 	atomic_t		mcg_attach;
396 
397 	spinlock_t		pending_lock; /* guard pending_mmaps */
398 	struct list_head	pending_mmaps;
399 
400 	spinlock_t		mmap_offset_lock; /* guard mmap_offset */
401 	u64			mmap_offset;
402 
403 	atomic64_t		stats_counters[RXE_NUM_OF_COUNTERS];
404 
405 	struct rxe_port		port;
406 	struct crypto_shash	*tfm;
407 };
408 
rxe_counter_inc(struct rxe_dev * rxe,enum rxe_counters index)409 static inline void rxe_counter_inc(struct rxe_dev *rxe, enum rxe_counters index)
410 {
411 	atomic64_inc(&rxe->stats_counters[index]);
412 }
413 
to_rdev(struct ib_device * dev)414 static inline struct rxe_dev *to_rdev(struct ib_device *dev)
415 {
416 	return dev ? container_of(dev, struct rxe_dev, ib_dev) : NULL;
417 }
418 
to_ruc(struct ib_ucontext * uc)419 static inline struct rxe_ucontext *to_ruc(struct ib_ucontext *uc)
420 {
421 	return uc ? container_of(uc, struct rxe_ucontext, ibuc) : NULL;
422 }
423 
to_rpd(struct ib_pd * pd)424 static inline struct rxe_pd *to_rpd(struct ib_pd *pd)
425 {
426 	return pd ? container_of(pd, struct rxe_pd, ibpd) : NULL;
427 }
428 
to_rah(struct ib_ah * ah)429 static inline struct rxe_ah *to_rah(struct ib_ah *ah)
430 {
431 	return ah ? container_of(ah, struct rxe_ah, ibah) : NULL;
432 }
433 
to_rsrq(struct ib_srq * srq)434 static inline struct rxe_srq *to_rsrq(struct ib_srq *srq)
435 {
436 	return srq ? container_of(srq, struct rxe_srq, ibsrq) : NULL;
437 }
438 
to_rqp(struct ib_qp * qp)439 static inline struct rxe_qp *to_rqp(struct ib_qp *qp)
440 {
441 	return qp ? container_of(qp, struct rxe_qp, ibqp) : NULL;
442 }
443 
to_rcq(struct ib_cq * cq)444 static inline struct rxe_cq *to_rcq(struct ib_cq *cq)
445 {
446 	return cq ? container_of(cq, struct rxe_cq, ibcq) : NULL;
447 }
448 
to_rmr(struct ib_mr * mr)449 static inline struct rxe_mr *to_rmr(struct ib_mr *mr)
450 {
451 	return mr ? container_of(mr, struct rxe_mr, ibmr) : NULL;
452 }
453 
to_rmw(struct ib_mw * mw)454 static inline struct rxe_mw *to_rmw(struct ib_mw *mw)
455 {
456 	return mw ? container_of(mw, struct rxe_mw, ibmw) : NULL;
457 }
458 
rxe_ah_pd(struct rxe_ah * ah)459 static inline struct rxe_pd *rxe_ah_pd(struct rxe_ah *ah)
460 {
461 	return to_rpd(ah->ibah.pd);
462 }
463 
mr_pd(struct rxe_mr * mr)464 static inline struct rxe_pd *mr_pd(struct rxe_mr *mr)
465 {
466 	return to_rpd(mr->ibmr.pd);
467 }
468 
rxe_mw_pd(struct rxe_mw * mw)469 static inline struct rxe_pd *rxe_mw_pd(struct rxe_mw *mw)
470 {
471 	return to_rpd(mw->ibmw.pd);
472 }
473 
474 int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name);
475 
476 #endif /* RXE_VERBS_H */
477