1 /* SPDX-License-Identifier: GPL-2.0-or-later */
2 /*
3  * RDMA Transport Layer
4  *
5  * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved.
6  * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved.
7  * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved.
8  */
9 
10 #ifndef RTRS_PRI_H
11 #define RTRS_PRI_H
12 
13 #include <linux/uuid.h>
14 #include <rdma/rdma_cm.h>
15 #include <rdma/ib_verbs.h>
16 #include <rdma/ib.h>
17 
18 #include "rtrs.h"
19 
20 #define RTRS_PROTO_VER_MAJOR 2
21 #define RTRS_PROTO_VER_MINOR 0
22 
23 #define RTRS_PROTO_VER_STRING __stringify(RTRS_PROTO_VER_MAJOR) "." \
24 			       __stringify(RTRS_PROTO_VER_MINOR)
25 
26 /*
27  * Max IB immediate data size is 2^28 (MAX_IMM_PAYL_BITS)
28  * and the minimum chunk size is 4096 (2^12).
29  * So the maximum sess_queue_depth is 65535 (2^16 - 1) in theory
30  * since queue_depth in rtrs_msg_conn_rsp is defined as le16.
31  * Therefore the pratical max value of sess_queue_depth is
32  * somewhere between 1 and 65535 and it depends on the system.
33  */
34 #define MAX_SESS_QUEUE_DEPTH 65535
35 
36 enum rtrs_imm_const {
37 	MAX_IMM_TYPE_BITS = 4,
38 	MAX_IMM_TYPE_MASK = ((1 << MAX_IMM_TYPE_BITS) - 1),
39 	MAX_IMM_PAYL_BITS = 28,
40 	MAX_IMM_PAYL_MASK = ((1 << MAX_IMM_PAYL_BITS) - 1),
41 };
42 
43 enum rtrs_imm_type {
44 	RTRS_IO_REQ_IMM       = 0, /* client to server */
45 	RTRS_IO_RSP_IMM       = 1, /* server to client */
46 	RTRS_IO_RSP_W_INV_IMM = 2, /* server to client */
47 
48 	RTRS_HB_MSG_IMM = 8, /* HB: HeartBeat */
49 	RTRS_HB_ACK_IMM = 9,
50 
51 	RTRS_LAST_IMM,
52 };
53 
54 enum {
55 	SERVICE_CON_QUEUE_DEPTH = 512,
56 
57 	MAX_PATHS_NUM = 128,
58 
59 	MIN_CHUNK_SIZE = 8192,
60 
61 	RTRS_HB_INTERVAL_MS = 5000,
62 	RTRS_HB_MISSED_MAX = 5,
63 
64 	RTRS_MAGIC = 0x1BBD,
65 	RTRS_PROTO_VER = (RTRS_PROTO_VER_MAJOR << 8) | RTRS_PROTO_VER_MINOR,
66 };
67 
68 struct rtrs_ib_dev;
69 
70 struct rtrs_rdma_dev_pd_ops {
71 	int (*init)(struct rtrs_ib_dev *dev);
72 };
73 
74 struct rtrs_rdma_dev_pd {
75 	struct mutex		mutex;
76 	struct list_head	list;
77 	enum ib_pd_flags	pd_flags;
78 	const struct rtrs_rdma_dev_pd_ops *ops;
79 };
80 
81 struct rtrs_ib_dev {
82 	struct ib_device	 *ib_dev;
83 	struct ib_pd		 *ib_pd;
84 	struct kref		 ref;
85 	struct list_head	 entry;
86 	struct rtrs_rdma_dev_pd *pool;
87 };
88 
89 struct rtrs_con {
90 	struct rtrs_path	*path;
91 	struct ib_qp		*qp;
92 	struct ib_cq		*cq;
93 	struct rdma_cm_id	*cm_id;
94 	unsigned int		cid;
95 	int                     nr_cqe;
96 	atomic_t		wr_cnt;
97 	atomic_t		sq_wr_avail;
98 };
99 
100 struct rtrs_path {
101 	struct list_head	entry;
102 	struct sockaddr_storage dst_addr;
103 	struct sockaddr_storage src_addr;
104 	char			sessname[NAME_MAX];
105 	uuid_t			uuid;
106 	struct rtrs_con	**con;
107 	unsigned int		con_num;
108 	unsigned int		irq_con_num;
109 	unsigned int		recon_cnt;
110 	unsigned int		signal_interval;
111 	struct rtrs_ib_dev	*dev;
112 	int			dev_ref;
113 	struct ib_cqe		*hb_cqe;
114 	void			(*hb_err_handler)(struct rtrs_con *con);
115 	struct workqueue_struct *hb_wq;
116 	struct delayed_work	hb_dwork;
117 	unsigned int		hb_interval_ms;
118 	unsigned int		hb_missed_cnt;
119 	unsigned int		hb_missed_max;
120 	ktime_t			hb_last_sent;
121 	ktime_t			hb_cur_latency;
122 };
123 
124 /* rtrs information unit */
125 struct rtrs_iu {
126 	struct ib_cqe           cqe;
127 	dma_addr_t              dma_addr;
128 	void                    *buf;
129 	size_t                  size;
130 	enum dma_data_direction direction;
131 };
132 
133 /**
134  * enum rtrs_msg_types - RTRS message types, see also rtrs/README
135  * @RTRS_MSG_INFO_REQ:		Client additional info request to the server
136  * @RTRS_MSG_INFO_RSP:		Server additional info response to the client
137  * @RTRS_MSG_WRITE:		Client writes data per RDMA to server
138  * @RTRS_MSG_READ:		Client requests data transfer from server
139  * @RTRS_MSG_RKEY_RSP:		Server refreshed rkey for rbuf
140  */
141 enum rtrs_msg_types {
142 	RTRS_MSG_INFO_REQ,
143 	RTRS_MSG_INFO_RSP,
144 	RTRS_MSG_WRITE,
145 	RTRS_MSG_READ,
146 	RTRS_MSG_RKEY_RSP,
147 };
148 
149 /**
150  * enum rtrs_msg_flags - RTRS message flags.
151  * @RTRS_NEED_INVAL:	Send invalidation in response.
152  * @RTRS_MSG_NEW_RKEY_F: Send refreshed rkey in response.
153  */
154 enum rtrs_msg_flags {
155 	RTRS_MSG_NEED_INVAL_F = 1 << 0,
156 	RTRS_MSG_NEW_RKEY_F = 1 << 1,
157 };
158 
159 /**
160  * struct rtrs_sg_desc - RDMA-Buffer entry description
161  * @addr:	Address of RDMA destination buffer
162  * @key:	Authorization rkey to write to the buffer
163  * @len:	Size of the buffer
164  */
165 struct rtrs_sg_desc {
166 	__le64			addr;
167 	__le32			key;
168 	__le32			len;
169 };
170 
171 /**
172  * struct rtrs_msg_conn_req - Client connection request to the server
173  * @magic:	   RTRS magic
174  * @version:	   RTRS protocol version
175  * @cid:	   Current connection id
176  * @cid_num:	   Number of connections per session
177  * @recon_cnt:	   Reconnections counter
178  * @sess_uuid:	   UUID of a session (path)
179  * @paths_uuid:	   UUID of a group of sessions (paths)
180  *
181  * NOTE: max size 56 bytes, see man rdma_connect().
182  */
183 struct rtrs_msg_conn_req {
184 	/* Is set to 0 by cma.c in case of AF_IB, do not touch that.
185 	 * see https://www.spinics.net/lists/linux-rdma/msg22397.html
186 	 */
187 	u8		__cma_version;
188 	/* On sender side that should be set to 0, or cma_save_ip_info()
189 	 * extract garbage and will fail.
190 	 */
191 	u8		__ip_version;
192 	__le16		magic;
193 	__le16		version;
194 	__le16		cid;
195 	__le16		cid_num;
196 	__le16		recon_cnt;
197 	uuid_t		sess_uuid;
198 	uuid_t		paths_uuid;
199 	u8		first_conn : 1;
200 	u8		reserved_bits : 7;
201 	u8		reserved[11];
202 };
203 
204 /**
205  * struct rtrs_msg_conn_rsp - Server connection response to the client
206  * @magic:	   RTRS magic
207  * @version:	   RTRS protocol version
208  * @errno:	   If rdma_accept() then 0, if rdma_reject() indicates error
209  * @queue_depth:   max inflight messages (queue-depth) in this session
210  * @max_io_size:   max io size server supports
211  * @max_hdr_size:  max msg header size server supports
212  *
213  * NOTE: size is 56 bytes, max possible is 136 bytes, see man rdma_accept().
214  */
215 struct rtrs_msg_conn_rsp {
216 	__le16		magic;
217 	__le16		version;
218 	__le16		errno;
219 	__le16		queue_depth;
220 	__le32		max_io_size;
221 	__le32		max_hdr_size;
222 	__le32		flags;
223 	u8		reserved[36];
224 };
225 
226 /**
227  * struct rtrs_msg_info_req
228  * @type:		@RTRS_MSG_INFO_REQ
229  * @pathname:		Path name chosen by client
230  */
231 struct rtrs_msg_info_req {
232 	__le16		type;
233 	u8		pathname[NAME_MAX];
234 	u8		reserved[15];
235 };
236 
237 /**
238  * struct rtrs_msg_info_rsp
239  * @type:		@RTRS_MSG_INFO_RSP
240  * @sg_cnt:		Number of @desc entries
241  * @desc:		RDMA buffers where the client can write to server
242  */
243 struct rtrs_msg_info_rsp {
244 	__le16		type;
245 	__le16          sg_cnt;
246 	u8              reserved[4];
247 	struct rtrs_sg_desc desc[];
248 };
249 
250 /**
251  * struct rtrs_msg_rkey_rsp
252  * @type:		@RTRS_MSG_RKEY_RSP
253  * @buf_id:		RDMA buf_id of the new rkey
254  * @rkey:		new remote key for RDMA buffers id from server
255  */
256 struct rtrs_msg_rkey_rsp {
257 	__le16		type;
258 	__le16          buf_id;
259 	__le32		rkey;
260 };
261 
262 /**
263  * struct rtrs_msg_rdma_read - RDMA data transfer request from client
264  * @type:		always @RTRS_MSG_READ
265  * @usr_len:		length of user payload
266  * @sg_cnt:		number of @desc entries
267  * @desc:		RDMA buffers where the server can write the result to
268  */
269 struct rtrs_msg_rdma_read {
270 	__le16			type;
271 	__le16			usr_len;
272 	__le16			flags;
273 	__le16			sg_cnt;
274 	struct rtrs_sg_desc    desc[];
275 };
276 
277 /**
278  * struct_msg_rdma_write - Message transferred to server with RDMA-Write
279  * @type:		always @RTRS_MSG_WRITE
280  * @usr_len:		length of user payload
281  */
282 struct rtrs_msg_rdma_write {
283 	__le16			type;
284 	__le16			usr_len;
285 };
286 
287 /**
288  * struct_msg_rdma_hdr - header for read or write request
289  * @type:		@RTRS_MSG_WRITE | @RTRS_MSG_READ
290  */
291 struct rtrs_msg_rdma_hdr {
292 	__le16			type;
293 };
294 
295 /* rtrs.c */
296 
297 struct rtrs_iu *rtrs_iu_alloc(u32 queue_num, size_t size, gfp_t t,
298 			      struct ib_device *dev, enum dma_data_direction,
299 			      void (*done)(struct ib_cq *cq, struct ib_wc *wc));
300 void rtrs_iu_free(struct rtrs_iu *iu, struct ib_device *dev, u32 queue_num);
301 int rtrs_iu_post_recv(struct rtrs_con *con, struct rtrs_iu *iu);
302 int rtrs_iu_post_send(struct rtrs_con *con, struct rtrs_iu *iu, size_t size,
303 		      struct ib_send_wr *head);
304 int rtrs_iu_post_rdma_write_imm(struct rtrs_con *con, struct rtrs_iu *iu,
305 				struct ib_sge *sge, unsigned int num_sge,
306 				u32 rkey, u64 rdma_addr, u32 imm_data,
307 				enum ib_send_flags flags,
308 				struct ib_send_wr *head,
309 				struct ib_send_wr *tail);
310 
311 int rtrs_post_recv_empty(struct rtrs_con *con, struct ib_cqe *cqe);
312 
313 int rtrs_cq_qp_create(struct rtrs_path *path, struct rtrs_con *con,
314 		      u32 max_send_sge, int cq_vector, int nr_cqe,
315 		      u32 max_send_wr, u32 max_recv_wr,
316 		      enum ib_poll_context poll_ctx);
317 void rtrs_cq_qp_destroy(struct rtrs_con *con);
318 
319 void rtrs_init_hb(struct rtrs_path *path, struct ib_cqe *cqe,
320 		  unsigned int interval_ms, unsigned int missed_max,
321 		  void (*err_handler)(struct rtrs_con *con),
322 		  struct workqueue_struct *wq);
323 void rtrs_start_hb(struct rtrs_path *path);
324 void rtrs_stop_hb(struct rtrs_path *path);
325 void rtrs_send_hb_ack(struct rtrs_path *path);
326 
327 void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags,
328 			   struct rtrs_rdma_dev_pd *pool);
329 void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool);
330 
331 struct rtrs_ib_dev *rtrs_ib_dev_find_or_add(struct ib_device *ib_dev,
332 					    struct rtrs_rdma_dev_pd *pool);
333 int rtrs_ib_dev_put(struct rtrs_ib_dev *dev);
334 
rtrs_to_imm(u32 type,u32 payload)335 static inline u32 rtrs_to_imm(u32 type, u32 payload)
336 {
337 	BUILD_BUG_ON(MAX_IMM_PAYL_BITS + MAX_IMM_TYPE_BITS != 32);
338 	BUILD_BUG_ON(RTRS_LAST_IMM > (1<<MAX_IMM_TYPE_BITS));
339 	return ((type & MAX_IMM_TYPE_MASK) << MAX_IMM_PAYL_BITS) |
340 		(payload & MAX_IMM_PAYL_MASK);
341 }
342 
rtrs_from_imm(u32 imm,u32 * type,u32 * payload)343 static inline void rtrs_from_imm(u32 imm, u32 *type, u32 *payload)
344 {
345 	*payload = imm & MAX_IMM_PAYL_MASK;
346 	*type = imm >> MAX_IMM_PAYL_BITS;
347 }
348 
rtrs_to_io_req_imm(u32 addr)349 static inline u32 rtrs_to_io_req_imm(u32 addr)
350 {
351 	return rtrs_to_imm(RTRS_IO_REQ_IMM, addr);
352 }
353 
rtrs_to_io_rsp_imm(u32 msg_id,int errno,bool w_inval)354 static inline u32 rtrs_to_io_rsp_imm(u32 msg_id, int errno, bool w_inval)
355 {
356 	enum rtrs_imm_type type;
357 	u32 payload;
358 
359 	/* 9 bits for errno, 19 bits for msg_id */
360 	payload = (abs(errno) & 0x1ff) << 19 | (msg_id & 0x7ffff);
361 	type = w_inval ? RTRS_IO_RSP_W_INV_IMM : RTRS_IO_RSP_IMM;
362 
363 	return rtrs_to_imm(type, payload);
364 }
365 
rtrs_from_io_rsp_imm(u32 payload,u32 * msg_id,int * errno)366 static inline void rtrs_from_io_rsp_imm(u32 payload, u32 *msg_id, int *errno)
367 {
368 	/* 9 bits for errno, 19 bits for msg_id */
369 	*msg_id = payload & 0x7ffff;
370 	*errno = -(int)((payload >> 19) & 0x1ff);
371 }
372 
373 #define STAT_STORE_FUNC(type, set_value, reset)				\
374 static ssize_t set_value##_store(struct kobject *kobj,			\
375 			     struct kobj_attribute *attr,		\
376 			     const char *buf, size_t count)		\
377 {									\
378 	int ret = -EINVAL;						\
379 	type *stats = container_of(kobj, type, kobj_stats);		\
380 									\
381 	if (sysfs_streq(buf, "1"))					\
382 		ret = reset(stats, true);			\
383 	else if (sysfs_streq(buf, "0"))					\
384 		ret = reset(stats, false);			\
385 	if (ret)							\
386 		return ret;						\
387 									\
388 	return count;							\
389 }
390 
391 #define STAT_SHOW_FUNC(type, get_value, print)				\
392 static ssize_t get_value##_show(struct kobject *kobj,			\
393 			   struct kobj_attribute *attr,			\
394 			   char *page)					\
395 {									\
396 	type *stats = container_of(kobj, type, kobj_stats);		\
397 									\
398 	return print(stats, page);			\
399 }
400 
401 #define STAT_ATTR(type, stat, print, reset)				\
402 STAT_STORE_FUNC(type, stat, reset)					\
403 STAT_SHOW_FUNC(type, stat, print)					\
404 static struct kobj_attribute stat##_attr = __ATTR_RW(stat)
405 
406 #endif /* RTRS_PRI_H */
407