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 enum rtrs_imm_const {
27 MAX_IMM_TYPE_BITS = 4,
28 MAX_IMM_TYPE_MASK = ((1 << MAX_IMM_TYPE_BITS) - 1),
29 MAX_IMM_PAYL_BITS = 28,
30 MAX_IMM_PAYL_MASK = ((1 << MAX_IMM_PAYL_BITS) - 1),
31 };
32
33 enum rtrs_imm_type {
34 RTRS_IO_REQ_IMM = 0, /* client to server */
35 RTRS_IO_RSP_IMM = 1, /* server to client */
36 RTRS_IO_RSP_W_INV_IMM = 2, /* server to client */
37
38 RTRS_HB_MSG_IMM = 8, /* HB: HeartBeat */
39 RTRS_HB_ACK_IMM = 9,
40
41 RTRS_LAST_IMM,
42 };
43
44 enum {
45 SERVICE_CON_QUEUE_DEPTH = 512,
46
47 MAX_PATHS_NUM = 128,
48
49 /*
50 * With the size of struct rtrs_permit allocated on the client, 4K
51 * is the maximum number of rtrs_permits we can allocate. This number is
52 * also used on the client to allocate the IU for the user connection
53 * to receive the RDMA addresses from the server.
54 */
55 MAX_SESS_QUEUE_DEPTH = 4096,
56
57 RTRS_HB_INTERVAL_MS = 5000,
58 RTRS_HB_MISSED_MAX = 5,
59
60 RTRS_MAGIC = 0x1BBD,
61 RTRS_PROTO_VER = (RTRS_PROTO_VER_MAJOR << 8) | RTRS_PROTO_VER_MINOR,
62 };
63
64 struct rtrs_ib_dev;
65
66 struct rtrs_rdma_dev_pd_ops {
67 struct rtrs_ib_dev *(*alloc)(void);
68 void (*free)(struct rtrs_ib_dev *dev);
69 int (*init)(struct rtrs_ib_dev *dev);
70 void (*deinit)(struct rtrs_ib_dev *dev);
71 };
72
73 struct rtrs_rdma_dev_pd {
74 struct mutex mutex;
75 struct list_head list;
76 enum ib_pd_flags pd_flags;
77 const struct rtrs_rdma_dev_pd_ops *ops;
78 };
79
80 struct rtrs_ib_dev {
81 struct ib_device *ib_dev;
82 struct ib_pd *ib_pd;
83 struct kref ref;
84 struct list_head entry;
85 struct rtrs_rdma_dev_pd *pool;
86 };
87
88 struct rtrs_con {
89 struct rtrs_sess *sess;
90 struct ib_qp *qp;
91 struct ib_cq *cq;
92 struct rdma_cm_id *cm_id;
93 unsigned int cid;
94 };
95
96 struct rtrs_sess {
97 struct list_head entry;
98 struct sockaddr_storage dst_addr;
99 struct sockaddr_storage src_addr;
100 char sessname[NAME_MAX];
101 uuid_t uuid;
102 struct rtrs_con **con;
103 unsigned int con_num;
104 unsigned int recon_cnt;
105 struct rtrs_ib_dev *dev;
106 int dev_ref;
107 struct ib_cqe *hb_cqe;
108 void (*hb_err_handler)(struct rtrs_con *con);
109 struct workqueue_struct *hb_wq;
110 struct delayed_work hb_dwork;
111 unsigned int hb_interval_ms;
112 unsigned int hb_missed_cnt;
113 unsigned int hb_missed_max;
114 };
115
116 /* rtrs information unit */
117 struct rtrs_iu {
118 struct ib_cqe cqe;
119 dma_addr_t dma_addr;
120 void *buf;
121 size_t size;
122 enum dma_data_direction direction;
123 };
124
125 /**
126 * enum rtrs_msg_types - RTRS message types, see also rtrs/README
127 * @RTRS_MSG_INFO_REQ: Client additional info request to the server
128 * @RTRS_MSG_INFO_RSP: Server additional info response to the client
129 * @RTRS_MSG_WRITE: Client writes data per RDMA to server
130 * @RTRS_MSG_READ: Client requests data transfer from server
131 * @RTRS_MSG_RKEY_RSP: Server refreshed rkey for rbuf
132 */
133 enum rtrs_msg_types {
134 RTRS_MSG_INFO_REQ,
135 RTRS_MSG_INFO_RSP,
136 RTRS_MSG_WRITE,
137 RTRS_MSG_READ,
138 RTRS_MSG_RKEY_RSP,
139 };
140
141 /**
142 * enum rtrs_msg_flags - RTRS message flags.
143 * @RTRS_NEED_INVAL: Send invalidation in response.
144 * @RTRS_MSG_NEW_RKEY_F: Send refreshed rkey in response.
145 */
146 enum rtrs_msg_flags {
147 RTRS_MSG_NEED_INVAL_F = 1 << 0,
148 RTRS_MSG_NEW_RKEY_F = 1 << 1,
149 };
150
151 /**
152 * struct rtrs_sg_desc - RDMA-Buffer entry description
153 * @addr: Address of RDMA destination buffer
154 * @key: Authorization rkey to write to the buffer
155 * @len: Size of the buffer
156 */
157 struct rtrs_sg_desc {
158 __le64 addr;
159 __le32 key;
160 __le32 len;
161 };
162
163 /**
164 * struct rtrs_msg_conn_req - Client connection request to the server
165 * @magic: RTRS magic
166 * @version: RTRS protocol version
167 * @cid: Current connection id
168 * @cid_num: Number of connections per session
169 * @recon_cnt: Reconnections counter
170 * @sess_uuid: UUID of a session (path)
171 * @paths_uuid: UUID of a group of sessions (paths)
172 *
173 * NOTE: max size 56 bytes, see man rdma_connect().
174 */
175 struct rtrs_msg_conn_req {
176 /* Is set to 0 by cma.c in case of AF_IB, do not touch that.
177 * see https://www.spinics.net/lists/linux-rdma/msg22397.html
178 */
179 u8 __cma_version;
180 /* On sender side that should be set to 0, or cma_save_ip_info()
181 * extract garbage and will fail.
182 */
183 u8 __ip_version;
184 __le16 magic;
185 __le16 version;
186 __le16 cid;
187 __le16 cid_num;
188 __le16 recon_cnt;
189 uuid_t sess_uuid;
190 uuid_t paths_uuid;
191 u8 reserved[12];
192 };
193
194 /**
195 * struct rtrs_msg_conn_rsp - Server connection response to the client
196 * @magic: RTRS magic
197 * @version: RTRS protocol version
198 * @errno: If rdma_accept() then 0, if rdma_reject() indicates error
199 * @queue_depth: max inflight messages (queue-depth) in this session
200 * @max_io_size: max io size server supports
201 * @max_hdr_size: max msg header size server supports
202 *
203 * NOTE: size is 56 bytes, max possible is 136 bytes, see man rdma_accept().
204 */
205 struct rtrs_msg_conn_rsp {
206 __le16 magic;
207 __le16 version;
208 __le16 errno;
209 __le16 queue_depth;
210 __le32 max_io_size;
211 __le32 max_hdr_size;
212 __le32 flags;
213 u8 reserved[36];
214 };
215
216 /**
217 * struct rtrs_msg_info_req
218 * @type: @RTRS_MSG_INFO_REQ
219 * @sessname: Session name chosen by client
220 */
221 struct rtrs_msg_info_req {
222 __le16 type;
223 u8 sessname[NAME_MAX];
224 u8 reserved[15];
225 };
226
227 /**
228 * struct rtrs_msg_info_rsp
229 * @type: @RTRS_MSG_INFO_RSP
230 * @sg_cnt: Number of @desc entries
231 * @desc: RDMA buffers where the client can write to server
232 */
233 struct rtrs_msg_info_rsp {
234 __le16 type;
235 __le16 sg_cnt;
236 u8 reserved[4];
237 struct rtrs_sg_desc desc[];
238 };
239
240 /**
241 * struct rtrs_msg_rkey_rsp
242 * @type: @RTRS_MSG_RKEY_RSP
243 * @buf_id: RDMA buf_id of the new rkey
244 * @rkey: new remote key for RDMA buffers id from server
245 */
246 struct rtrs_msg_rkey_rsp {
247 __le16 type;
248 __le16 buf_id;
249 __le32 rkey;
250 };
251
252 /**
253 * struct rtrs_msg_rdma_read - RDMA data transfer request from client
254 * @type: always @RTRS_MSG_READ
255 * @usr_len: length of user payload
256 * @sg_cnt: number of @desc entries
257 * @desc: RDMA buffers where the server can write the result to
258 */
259 struct rtrs_msg_rdma_read {
260 __le16 type;
261 __le16 usr_len;
262 __le16 flags;
263 __le16 sg_cnt;
264 struct rtrs_sg_desc desc[];
265 };
266
267 /**
268 * struct_msg_rdma_write - Message transferred to server with RDMA-Write
269 * @type: always @RTRS_MSG_WRITE
270 * @usr_len: length of user payload
271 */
272 struct rtrs_msg_rdma_write {
273 __le16 type;
274 __le16 usr_len;
275 };
276
277 /**
278 * struct_msg_rdma_hdr - header for read or write request
279 * @type: @RTRS_MSG_WRITE | @RTRS_MSG_READ
280 */
281 struct rtrs_msg_rdma_hdr {
282 __le16 type;
283 };
284
285 /* rtrs.c */
286
287 struct rtrs_iu *rtrs_iu_alloc(u32 queue_size, size_t size, gfp_t t,
288 struct ib_device *dev, enum dma_data_direction,
289 void (*done)(struct ib_cq *cq, struct ib_wc *wc));
290 void rtrs_iu_free(struct rtrs_iu *iu, enum dma_data_direction dir,
291 struct ib_device *dev, u32 queue_size);
292 int rtrs_iu_post_recv(struct rtrs_con *con, struct rtrs_iu *iu);
293 int rtrs_iu_post_send(struct rtrs_con *con, struct rtrs_iu *iu, size_t size,
294 struct ib_send_wr *head);
295 int rtrs_iu_post_rdma_write_imm(struct rtrs_con *con, struct rtrs_iu *iu,
296 struct ib_sge *sge, unsigned int num_sge,
297 u32 rkey, u64 rdma_addr, u32 imm_data,
298 enum ib_send_flags flags,
299 struct ib_send_wr *head);
300
301 int rtrs_post_recv_empty(struct rtrs_con *con, struct ib_cqe *cqe);
302 int rtrs_post_rdma_write_imm_empty(struct rtrs_con *con, struct ib_cqe *cqe,
303 u32 imm_data, enum ib_send_flags flags,
304 struct ib_send_wr *head);
305
306 int rtrs_cq_qp_create(struct rtrs_sess *rtrs_sess, struct rtrs_con *con,
307 u32 max_send_sge, int cq_vector, u16 cq_size,
308 u16 wr_queue_size, enum ib_poll_context poll_ctx);
309 void rtrs_cq_qp_destroy(struct rtrs_con *con);
310
311 void rtrs_init_hb(struct rtrs_sess *sess, struct ib_cqe *cqe,
312 unsigned int interval_ms, unsigned int missed_max,
313 void (*err_handler)(struct rtrs_con *con),
314 struct workqueue_struct *wq);
315 void rtrs_start_hb(struct rtrs_sess *sess);
316 void rtrs_stop_hb(struct rtrs_sess *sess);
317 void rtrs_send_hb_ack(struct rtrs_sess *sess);
318
319 void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags,
320 struct rtrs_rdma_dev_pd *pool);
321 void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool);
322
323 struct rtrs_ib_dev *rtrs_ib_dev_find_or_add(struct ib_device *ib_dev,
324 struct rtrs_rdma_dev_pd *pool);
325 int rtrs_ib_dev_put(struct rtrs_ib_dev *dev);
326
rtrs_to_imm(u32 type,u32 payload)327 static inline u32 rtrs_to_imm(u32 type, u32 payload)
328 {
329 BUILD_BUG_ON(MAX_IMM_PAYL_BITS + MAX_IMM_TYPE_BITS != 32);
330 BUILD_BUG_ON(RTRS_LAST_IMM > (1<<MAX_IMM_TYPE_BITS));
331 return ((type & MAX_IMM_TYPE_MASK) << MAX_IMM_PAYL_BITS) |
332 (payload & MAX_IMM_PAYL_MASK);
333 }
334
rtrs_from_imm(u32 imm,u32 * type,u32 * payload)335 static inline void rtrs_from_imm(u32 imm, u32 *type, u32 *payload)
336 {
337 *payload = imm & MAX_IMM_PAYL_MASK;
338 *type = imm >> MAX_IMM_PAYL_BITS;
339 }
340
rtrs_to_io_req_imm(u32 addr)341 static inline u32 rtrs_to_io_req_imm(u32 addr)
342 {
343 return rtrs_to_imm(RTRS_IO_REQ_IMM, addr);
344 }
345
rtrs_to_io_rsp_imm(u32 msg_id,int errno,bool w_inval)346 static inline u32 rtrs_to_io_rsp_imm(u32 msg_id, int errno, bool w_inval)
347 {
348 enum rtrs_imm_type type;
349 u32 payload;
350
351 /* 9 bits for errno, 19 bits for msg_id */
352 payload = (abs(errno) & 0x1ff) << 19 | (msg_id & 0x7ffff);
353 type = w_inval ? RTRS_IO_RSP_W_INV_IMM : RTRS_IO_RSP_IMM;
354
355 return rtrs_to_imm(type, payload);
356 }
357
rtrs_from_io_rsp_imm(u32 payload,u32 * msg_id,int * errno)358 static inline void rtrs_from_io_rsp_imm(u32 payload, u32 *msg_id, int *errno)
359 {
360 /* 9 bits for errno, 19 bits for msg_id */
361 *msg_id = payload & 0x7ffff;
362 *errno = -(int)((payload >> 19) & 0x1ff);
363 }
364
365 #define STAT_STORE_FUNC(type, set_value, reset) \
366 static ssize_t set_value##_store(struct kobject *kobj, \
367 struct kobj_attribute *attr, \
368 const char *buf, size_t count) \
369 { \
370 int ret = -EINVAL; \
371 type *stats = container_of(kobj, type, kobj_stats); \
372 \
373 if (sysfs_streq(buf, "1")) \
374 ret = reset(stats, true); \
375 else if (sysfs_streq(buf, "0")) \
376 ret = reset(stats, false); \
377 if (ret) \
378 return ret; \
379 \
380 return count; \
381 }
382
383 #define STAT_SHOW_FUNC(type, get_value, print) \
384 static ssize_t get_value##_show(struct kobject *kobj, \
385 struct kobj_attribute *attr, \
386 char *page) \
387 { \
388 type *stats = container_of(kobj, type, kobj_stats); \
389 \
390 return print(stats, page, PAGE_SIZE); \
391 }
392
393 #define STAT_ATTR(type, stat, print, reset) \
394 STAT_STORE_FUNC(type, stat, reset) \
395 STAT_SHOW_FUNC(type, stat, print) \
396 static struct kobj_attribute stat##_attr = __ATTR_RW(stat)
397
398 #endif /* RTRS_PRI_H */
399