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 65536 (2^16) in theory.
30 * But mempool_create, create_qp and ib_post_send fail with
31 * "cannot allocate memory" error if sess_queue_depth is too big.
32 * Therefore the pratical max value of sess_queue_depth is
33 * somewhere between 1 and 65534 and it depends on the system.
34 */
35 #define MAX_SESS_QUEUE_DEPTH 65535
36
37 enum rtrs_imm_const {
38 MAX_IMM_TYPE_BITS = 4,
39 MAX_IMM_TYPE_MASK = ((1 << MAX_IMM_TYPE_BITS) - 1),
40 MAX_IMM_PAYL_BITS = 28,
41 MAX_IMM_PAYL_MASK = ((1 << MAX_IMM_PAYL_BITS) - 1),
42 };
43
44 enum rtrs_imm_type {
45 RTRS_IO_REQ_IMM = 0, /* client to server */
46 RTRS_IO_RSP_IMM = 1, /* server to client */
47 RTRS_IO_RSP_W_INV_IMM = 2, /* server to client */
48
49 RTRS_HB_MSG_IMM = 8, /* HB: HeartBeat */
50 RTRS_HB_ACK_IMM = 9,
51
52 RTRS_LAST_IMM,
53 };
54
55 enum {
56 SERVICE_CON_QUEUE_DEPTH = 512,
57
58 MAX_PATHS_NUM = 128,
59
60 MIN_CHUNK_SIZE = 8192,
61
62 RTRS_HB_INTERVAL_MS = 5000,
63 RTRS_HB_MISSED_MAX = 5,
64
65 RTRS_MAGIC = 0x1BBD,
66 RTRS_PROTO_VER = (RTRS_PROTO_VER_MAJOR << 8) | RTRS_PROTO_VER_MINOR,
67 };
68
69 struct rtrs_ib_dev;
70
71 struct rtrs_rdma_dev_pd_ops {
72 struct rtrs_ib_dev *(*alloc)(void);
73 void (*free)(struct rtrs_ib_dev *dev);
74 int (*init)(struct rtrs_ib_dev *dev);
75 void (*deinit)(struct rtrs_ib_dev *dev);
76 };
77
78 struct rtrs_rdma_dev_pd {
79 struct mutex mutex;
80 struct list_head list;
81 enum ib_pd_flags pd_flags;
82 const struct rtrs_rdma_dev_pd_ops *ops;
83 };
84
85 struct rtrs_ib_dev {
86 struct ib_device *ib_dev;
87 struct ib_pd *ib_pd;
88 struct kref ref;
89 struct list_head entry;
90 struct rtrs_rdma_dev_pd *pool;
91 };
92
93 struct rtrs_con {
94 struct rtrs_sess *sess;
95 struct ib_qp *qp;
96 struct ib_cq *cq;
97 struct rdma_cm_id *cm_id;
98 unsigned int cid;
99 };
100
101 struct rtrs_sess {
102 struct list_head entry;
103 struct sockaddr_storage dst_addr;
104 struct sockaddr_storage src_addr;
105 char sessname[NAME_MAX];
106 uuid_t uuid;
107 struct rtrs_con **con;
108 unsigned int con_num;
109 unsigned int recon_cnt;
110 struct rtrs_ib_dev *dev;
111 int dev_ref;
112 struct ib_cqe *hb_cqe;
113 void (*hb_err_handler)(struct rtrs_con *con);
114 struct workqueue_struct *hb_wq;
115 struct delayed_work hb_dwork;
116 unsigned int hb_interval_ms;
117 unsigned int hb_missed_cnt;
118 unsigned int hb_missed_max;
119 };
120
121 /* rtrs information unit */
122 struct rtrs_iu {
123 struct ib_cqe cqe;
124 dma_addr_t dma_addr;
125 void *buf;
126 size_t size;
127 enum dma_data_direction direction;
128 };
129
130 /**
131 * enum rtrs_msg_types - RTRS message types, see also rtrs/README
132 * @RTRS_MSG_INFO_REQ: Client additional info request to the server
133 * @RTRS_MSG_INFO_RSP: Server additional info response to the client
134 * @RTRS_MSG_WRITE: Client writes data per RDMA to server
135 * @RTRS_MSG_READ: Client requests data transfer from server
136 * @RTRS_MSG_RKEY_RSP: Server refreshed rkey for rbuf
137 */
138 enum rtrs_msg_types {
139 RTRS_MSG_INFO_REQ,
140 RTRS_MSG_INFO_RSP,
141 RTRS_MSG_WRITE,
142 RTRS_MSG_READ,
143 RTRS_MSG_RKEY_RSP,
144 };
145
146 /**
147 * enum rtrs_msg_flags - RTRS message flags.
148 * @RTRS_NEED_INVAL: Send invalidation in response.
149 * @RTRS_MSG_NEW_RKEY_F: Send refreshed rkey in response.
150 */
151 enum rtrs_msg_flags {
152 RTRS_MSG_NEED_INVAL_F = 1 << 0,
153 RTRS_MSG_NEW_RKEY_F = 1 << 1,
154 };
155
156 /**
157 * struct rtrs_sg_desc - RDMA-Buffer entry description
158 * @addr: Address of RDMA destination buffer
159 * @key: Authorization rkey to write to the buffer
160 * @len: Size of the buffer
161 */
162 struct rtrs_sg_desc {
163 __le64 addr;
164 __le32 key;
165 __le32 len;
166 };
167
168 /**
169 * struct rtrs_msg_conn_req - Client connection request to the server
170 * @magic: RTRS magic
171 * @version: RTRS protocol version
172 * @cid: Current connection id
173 * @cid_num: Number of connections per session
174 * @recon_cnt: Reconnections counter
175 * @sess_uuid: UUID of a session (path)
176 * @paths_uuid: UUID of a group of sessions (paths)
177 *
178 * NOTE: max size 56 bytes, see man rdma_connect().
179 */
180 struct rtrs_msg_conn_req {
181 /* Is set to 0 by cma.c in case of AF_IB, do not touch that.
182 * see https://www.spinics.net/lists/linux-rdma/msg22397.html
183 */
184 u8 __cma_version;
185 /* On sender side that should be set to 0, or cma_save_ip_info()
186 * extract garbage and will fail.
187 */
188 u8 __ip_version;
189 __le16 magic;
190 __le16 version;
191 __le16 cid;
192 __le16 cid_num;
193 __le16 recon_cnt;
194 uuid_t sess_uuid;
195 uuid_t paths_uuid;
196 u8 first_conn : 1;
197 u8 reserved_bits : 7;
198 u8 reserved[11];
199 };
200
201 /**
202 * struct rtrs_msg_conn_rsp - Server connection response to the client
203 * @magic: RTRS magic
204 * @version: RTRS protocol version
205 * @errno: If rdma_accept() then 0, if rdma_reject() indicates error
206 * @queue_depth: max inflight messages (queue-depth) in this session
207 * @max_io_size: max io size server supports
208 * @max_hdr_size: max msg header size server supports
209 *
210 * NOTE: size is 56 bytes, max possible is 136 bytes, see man rdma_accept().
211 */
212 struct rtrs_msg_conn_rsp {
213 __le16 magic;
214 __le16 version;
215 __le16 errno;
216 __le16 queue_depth;
217 __le32 max_io_size;
218 __le32 max_hdr_size;
219 __le32 flags;
220 u8 reserved[36];
221 };
222
223 /**
224 * struct rtrs_msg_info_req
225 * @type: @RTRS_MSG_INFO_REQ
226 * @sessname: Session name chosen by client
227 */
228 struct rtrs_msg_info_req {
229 __le16 type;
230 u8 sessname[NAME_MAX];
231 u8 reserved[15];
232 };
233
234 /**
235 * struct rtrs_msg_info_rsp
236 * @type: @RTRS_MSG_INFO_RSP
237 * @sg_cnt: Number of @desc entries
238 * @desc: RDMA buffers where the client can write to server
239 */
240 struct rtrs_msg_info_rsp {
241 __le16 type;
242 __le16 sg_cnt;
243 u8 reserved[4];
244 struct rtrs_sg_desc desc[];
245 };
246
247 /**
248 * struct rtrs_msg_rkey_rsp
249 * @type: @RTRS_MSG_RKEY_RSP
250 * @buf_id: RDMA buf_id of the new rkey
251 * @rkey: new remote key for RDMA buffers id from server
252 */
253 struct rtrs_msg_rkey_rsp {
254 __le16 type;
255 __le16 buf_id;
256 __le32 rkey;
257 };
258
259 /**
260 * struct rtrs_msg_rdma_read - RDMA data transfer request from client
261 * @type: always @RTRS_MSG_READ
262 * @usr_len: length of user payload
263 * @sg_cnt: number of @desc entries
264 * @desc: RDMA buffers where the server can write the result to
265 */
266 struct rtrs_msg_rdma_read {
267 __le16 type;
268 __le16 usr_len;
269 __le16 flags;
270 __le16 sg_cnt;
271 struct rtrs_sg_desc desc[];
272 };
273
274 /**
275 * struct_msg_rdma_write - Message transferred to server with RDMA-Write
276 * @type: always @RTRS_MSG_WRITE
277 * @usr_len: length of user payload
278 */
279 struct rtrs_msg_rdma_write {
280 __le16 type;
281 __le16 usr_len;
282 };
283
284 /**
285 * struct_msg_rdma_hdr - header for read or write request
286 * @type: @RTRS_MSG_WRITE | @RTRS_MSG_READ
287 */
288 struct rtrs_msg_rdma_hdr {
289 __le16 type;
290 };
291
292 /* rtrs.c */
293
294 struct rtrs_iu *rtrs_iu_alloc(u32 queue_size, size_t size, gfp_t t,
295 struct ib_device *dev, enum dma_data_direction,
296 void (*done)(struct ib_cq *cq, struct ib_wc *wc));
297 void rtrs_iu_free(struct rtrs_iu *iu, struct ib_device *dev, u32 queue_size);
298 int rtrs_iu_post_recv(struct rtrs_con *con, struct rtrs_iu *iu);
299 int rtrs_iu_post_send(struct rtrs_con *con, struct rtrs_iu *iu, size_t size,
300 struct ib_send_wr *head);
301 int rtrs_iu_post_rdma_write_imm(struct rtrs_con *con, struct rtrs_iu *iu,
302 struct ib_sge *sge, unsigned int num_sge,
303 u32 rkey, u64 rdma_addr, u32 imm_data,
304 enum ib_send_flags flags,
305 struct ib_send_wr *head);
306
307 int rtrs_post_recv_empty(struct rtrs_con *con, struct ib_cqe *cqe);
308 int rtrs_post_rdma_write_imm_empty(struct rtrs_con *con, struct ib_cqe *cqe,
309 u32 imm_data, enum ib_send_flags flags,
310 struct ib_send_wr *head);
311
312 int rtrs_cq_qp_create(struct rtrs_sess *rtrs_sess, struct rtrs_con *con,
313 u32 max_send_sge, int cq_vector, int cq_size,
314 u32 max_send_wr, u32 max_recv_wr,
315 enum ib_poll_context poll_ctx);
316 void rtrs_cq_qp_destroy(struct rtrs_con *con);
317
318 void rtrs_init_hb(struct rtrs_sess *sess, struct ib_cqe *cqe,
319 unsigned int interval_ms, unsigned int missed_max,
320 void (*err_handler)(struct rtrs_con *con),
321 struct workqueue_struct *wq);
322 void rtrs_start_hb(struct rtrs_sess *sess);
323 void rtrs_stop_hb(struct rtrs_sess *sess);
324 void rtrs_send_hb_ack(struct rtrs_sess *sess);
325
326 void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags,
327 struct rtrs_rdma_dev_pd *pool);
328 void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool);
329
330 struct rtrs_ib_dev *rtrs_ib_dev_find_or_add(struct ib_device *ib_dev,
331 struct rtrs_rdma_dev_pd *pool);
332 int rtrs_ib_dev_put(struct rtrs_ib_dev *dev);
333
rtrs_to_imm(u32 type,u32 payload)334 static inline u32 rtrs_to_imm(u32 type, u32 payload)
335 {
336 BUILD_BUG_ON(MAX_IMM_PAYL_BITS + MAX_IMM_TYPE_BITS != 32);
337 BUILD_BUG_ON(RTRS_LAST_IMM > (1<<MAX_IMM_TYPE_BITS));
338 return ((type & MAX_IMM_TYPE_MASK) << MAX_IMM_PAYL_BITS) |
339 (payload & MAX_IMM_PAYL_MASK);
340 }
341
rtrs_from_imm(u32 imm,u32 * type,u32 * payload)342 static inline void rtrs_from_imm(u32 imm, u32 *type, u32 *payload)
343 {
344 *payload = imm & MAX_IMM_PAYL_MASK;
345 *type = imm >> MAX_IMM_PAYL_BITS;
346 }
347
rtrs_to_io_req_imm(u32 addr)348 static inline u32 rtrs_to_io_req_imm(u32 addr)
349 {
350 return rtrs_to_imm(RTRS_IO_REQ_IMM, addr);
351 }
352
rtrs_to_io_rsp_imm(u32 msg_id,int errno,bool w_inval)353 static inline u32 rtrs_to_io_rsp_imm(u32 msg_id, int errno, bool w_inval)
354 {
355 enum rtrs_imm_type type;
356 u32 payload;
357
358 /* 9 bits for errno, 19 bits for msg_id */
359 payload = (abs(errno) & 0x1ff) << 19 | (msg_id & 0x7ffff);
360 type = w_inval ? RTRS_IO_RSP_W_INV_IMM : RTRS_IO_RSP_IMM;
361
362 return rtrs_to_imm(type, payload);
363 }
364
rtrs_from_io_rsp_imm(u32 payload,u32 * msg_id,int * errno)365 static inline void rtrs_from_io_rsp_imm(u32 payload, u32 *msg_id, int *errno)
366 {
367 /* 9 bits for errno, 19 bits for msg_id */
368 *msg_id = payload & 0x7ffff;
369 *errno = -(int)((payload >> 19) & 0x1ff);
370 }
371
372 #define STAT_STORE_FUNC(type, set_value, reset) \
373 static ssize_t set_value##_store(struct kobject *kobj, \
374 struct kobj_attribute *attr, \
375 const char *buf, size_t count) \
376 { \
377 int ret = -EINVAL; \
378 type *stats = container_of(kobj, type, kobj_stats); \
379 \
380 if (sysfs_streq(buf, "1")) \
381 ret = reset(stats, true); \
382 else if (sysfs_streq(buf, "0")) \
383 ret = reset(stats, false); \
384 if (ret) \
385 return ret; \
386 \
387 return count; \
388 }
389
390 #define STAT_SHOW_FUNC(type, get_value, print) \
391 static ssize_t get_value##_show(struct kobject *kobj, \
392 struct kobj_attribute *attr, \
393 char *page) \
394 { \
395 type *stats = container_of(kobj, type, kobj_stats); \
396 \
397 return print(stats, page, PAGE_SIZE); \
398 }
399
400 #define STAT_ATTR(type, stat, print, reset) \
401 STAT_STORE_FUNC(type, stat, reset) \
402 STAT_SHOW_FUNC(type, stat, print) \
403 static struct kobj_attribute stat##_attr = __ATTR_RW(stat)
404
405 #endif /* RTRS_PRI_H */
406