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 void (*deinit)(struct rtrs_ib_dev *dev);
73 };
74
75 struct rtrs_rdma_dev_pd {
76 struct mutex mutex;
77 struct list_head list;
78 enum ib_pd_flags pd_flags;
79 const struct rtrs_rdma_dev_pd_ops *ops;
80 };
81
82 struct rtrs_ib_dev {
83 struct ib_device *ib_dev;
84 struct ib_pd *ib_pd;
85 struct kref ref;
86 struct list_head entry;
87 struct rtrs_rdma_dev_pd *pool;
88 struct ib_event_handler event_handler;
89 };
90
91 struct rtrs_con {
92 struct rtrs_path *path;
93 struct ib_qp *qp;
94 struct ib_cq *cq;
95 struct rdma_cm_id *cm_id;
96 unsigned int cid;
97 int nr_cqe;
98 atomic_t wr_cnt;
99 atomic_t sq_wr_avail;
100 };
101
102 struct rtrs_path {
103 struct list_head entry;
104 struct sockaddr_storage dst_addr;
105 struct sockaddr_storage src_addr;
106 char sessname[NAME_MAX];
107 uuid_t uuid;
108 struct rtrs_con **con;
109 unsigned int con_num;
110 unsigned int irq_con_num;
111 unsigned int recon_cnt;
112 unsigned int signal_interval;
113 struct rtrs_ib_dev *dev;
114 int dev_ref;
115 struct ib_cqe *hb_cqe;
116 void (*hb_err_handler)(struct rtrs_con *con);
117 struct workqueue_struct *hb_wq;
118 struct delayed_work hb_dwork;
119 unsigned int hb_interval_ms;
120 unsigned int hb_missed_cnt;
121 unsigned int hb_missed_max;
122 ktime_t hb_last_sent;
123 ktime_t hb_cur_latency;
124 };
125
126 /* rtrs information unit */
127 struct rtrs_iu {
128 struct ib_cqe cqe;
129 dma_addr_t dma_addr;
130 void *buf;
131 size_t size;
132 enum dma_data_direction direction;
133 };
134
135 /**
136 * enum rtrs_msg_types - RTRS message types, see also rtrs/README
137 * @RTRS_MSG_INFO_REQ: Client additional info request to the server
138 * @RTRS_MSG_INFO_RSP: Server additional info response to the client
139 * @RTRS_MSG_WRITE: Client writes data per RDMA to server
140 * @RTRS_MSG_READ: Client requests data transfer from server
141 * @RTRS_MSG_RKEY_RSP: Server refreshed rkey for rbuf
142 */
143 enum rtrs_msg_types {
144 RTRS_MSG_INFO_REQ,
145 RTRS_MSG_INFO_RSP,
146 RTRS_MSG_WRITE,
147 RTRS_MSG_READ,
148 RTRS_MSG_RKEY_RSP,
149 };
150
151 /**
152 * enum rtrs_msg_flags - RTRS message flags.
153 * @RTRS_NEED_INVAL: Send invalidation in response.
154 * @RTRS_MSG_NEW_RKEY_F: Send refreshed rkey in response.
155 */
156 enum rtrs_msg_flags {
157 RTRS_MSG_NEED_INVAL_F = 1 << 0,
158 RTRS_MSG_NEW_RKEY_F = 1 << 1,
159 };
160
161 /**
162 * struct rtrs_sg_desc - RDMA-Buffer entry description
163 * @addr: Address of RDMA destination buffer
164 * @key: Authorization rkey to write to the buffer
165 * @len: Size of the buffer
166 */
167 struct rtrs_sg_desc {
168 __le64 addr;
169 __le32 key;
170 __le32 len;
171 };
172
173 /**
174 * struct rtrs_msg_conn_req - Client connection request to the server
175 * @magic: RTRS magic
176 * @version: RTRS protocol version
177 * @cid: Current connection id
178 * @cid_num: Number of connections per session
179 * @recon_cnt: Reconnections counter
180 * @sess_uuid: UUID of a session (path)
181 * @paths_uuid: UUID of a group of sessions (paths)
182 *
183 * NOTE: max size 56 bytes, see man rdma_connect().
184 */
185 struct rtrs_msg_conn_req {
186 /* Is set to 0 by cma.c in case of AF_IB, do not touch that.
187 * see https://www.spinics.net/lists/linux-rdma/msg22397.html
188 */
189 u8 __cma_version;
190 /* On sender side that should be set to 0, or cma_save_ip_info()
191 * extract garbage and will fail.
192 */
193 u8 __ip_version;
194 __le16 magic;
195 __le16 version;
196 __le16 cid;
197 __le16 cid_num;
198 __le16 recon_cnt;
199 uuid_t sess_uuid;
200 uuid_t paths_uuid;
201 u8 first_conn : 1;
202 u8 reserved_bits : 7;
203 u8 reserved[11];
204 };
205
206 /**
207 * struct rtrs_msg_conn_rsp - Server connection response to the client
208 * @magic: RTRS magic
209 * @version: RTRS protocol version
210 * @errno: If rdma_accept() then 0, if rdma_reject() indicates error
211 * @queue_depth: max inflight messages (queue-depth) in this session
212 * @max_io_size: max io size server supports
213 * @max_hdr_size: max msg header size server supports
214 *
215 * NOTE: size is 56 bytes, max possible is 136 bytes, see man rdma_accept().
216 */
217 struct rtrs_msg_conn_rsp {
218 __le16 magic;
219 __le16 version;
220 __le16 errno;
221 __le16 queue_depth;
222 __le32 max_io_size;
223 __le32 max_hdr_size;
224 __le32 flags;
225 u8 reserved[36];
226 };
227
228 /**
229 * struct rtrs_msg_info_req
230 * @type: @RTRS_MSG_INFO_REQ
231 * @pathname: Path name chosen by client
232 */
233 struct rtrs_msg_info_req {
234 __le16 type;
235 u8 pathname[NAME_MAX];
236 u8 reserved[15];
237 };
238
239 /**
240 * struct rtrs_msg_info_rsp
241 * @type: @RTRS_MSG_INFO_RSP
242 * @sg_cnt: Number of @desc entries
243 * @desc: RDMA buffers where the client can write to server
244 */
245 struct rtrs_msg_info_rsp {
246 __le16 type;
247 __le16 sg_cnt;
248 u8 reserved[4];
249 struct rtrs_sg_desc desc[];
250 };
251
252 /**
253 * struct rtrs_msg_rkey_rsp
254 * @type: @RTRS_MSG_RKEY_RSP
255 * @buf_id: RDMA buf_id of the new rkey
256 * @rkey: new remote key for RDMA buffers id from server
257 */
258 struct rtrs_msg_rkey_rsp {
259 __le16 type;
260 __le16 buf_id;
261 __le32 rkey;
262 };
263
264 /**
265 * struct rtrs_msg_rdma_read - RDMA data transfer request from client
266 * @type: always @RTRS_MSG_READ
267 * @usr_len: length of user payload
268 * @sg_cnt: number of @desc entries
269 * @desc: RDMA buffers where the server can write the result to
270 */
271 struct rtrs_msg_rdma_read {
272 __le16 type;
273 __le16 usr_len;
274 __le16 flags;
275 __le16 sg_cnt;
276 struct rtrs_sg_desc desc[];
277 };
278
279 /**
280 * struct_msg_rdma_write - Message transferred to server with RDMA-Write
281 * @type: always @RTRS_MSG_WRITE
282 * @usr_len: length of user payload
283 */
284 struct rtrs_msg_rdma_write {
285 __le16 type;
286 __le16 usr_len;
287 };
288
289 /**
290 * struct_msg_rdma_hdr - header for read or write request
291 * @type: @RTRS_MSG_WRITE | @RTRS_MSG_READ
292 */
293 struct rtrs_msg_rdma_hdr {
294 __le16 type;
295 };
296
297 /* rtrs.c */
298
299 struct rtrs_iu *rtrs_iu_alloc(u32 queue_num, size_t size, gfp_t t,
300 struct ib_device *dev, enum dma_data_direction,
301 void (*done)(struct ib_cq *cq, struct ib_wc *wc));
302 void rtrs_iu_free(struct rtrs_iu *iu, struct ib_device *dev, u32 queue_num);
303 int rtrs_iu_post_recv(struct rtrs_con *con, struct rtrs_iu *iu);
304 int rtrs_iu_post_send(struct rtrs_con *con, struct rtrs_iu *iu, size_t size,
305 struct ib_send_wr *head);
306 int rtrs_iu_post_rdma_write_imm(struct rtrs_con *con, struct rtrs_iu *iu,
307 struct ib_sge *sge, unsigned int num_sge,
308 u32 rkey, u64 rdma_addr, u32 imm_data,
309 enum ib_send_flags flags,
310 struct ib_send_wr *head,
311 struct ib_send_wr *tail);
312
313 int rtrs_post_recv_empty(struct rtrs_con *con, struct ib_cqe *cqe);
314
315 int rtrs_cq_qp_create(struct rtrs_path *path, struct rtrs_con *con,
316 u32 max_send_sge, int cq_vector, int nr_cqe,
317 u32 max_send_wr, u32 max_recv_wr,
318 enum ib_poll_context poll_ctx);
319 void rtrs_cq_qp_destroy(struct rtrs_con *con);
320
321 void rtrs_init_hb(struct rtrs_path *path, struct ib_cqe *cqe,
322 unsigned int interval_ms, unsigned int missed_max,
323 void (*err_handler)(struct rtrs_con *con),
324 struct workqueue_struct *wq);
325 void rtrs_start_hb(struct rtrs_path *path);
326 void rtrs_stop_hb(struct rtrs_path *path);
327 void rtrs_send_hb_ack(struct rtrs_path *path);
328
329 void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags,
330 struct rtrs_rdma_dev_pd *pool);
331 void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool);
332
333 struct rtrs_ib_dev *rtrs_ib_dev_find_or_add(struct ib_device *ib_dev,
334 struct rtrs_rdma_dev_pd *pool);
335 int rtrs_ib_dev_put(struct rtrs_ib_dev *dev);
336
rtrs_to_imm(u32 type,u32 payload)337 static inline u32 rtrs_to_imm(u32 type, u32 payload)
338 {
339 BUILD_BUG_ON(MAX_IMM_PAYL_BITS + MAX_IMM_TYPE_BITS != 32);
340 BUILD_BUG_ON(RTRS_LAST_IMM > (1<<MAX_IMM_TYPE_BITS));
341 return ((type & MAX_IMM_TYPE_MASK) << MAX_IMM_PAYL_BITS) |
342 (payload & MAX_IMM_PAYL_MASK);
343 }
344
rtrs_from_imm(u32 imm,u32 * type,u32 * payload)345 static inline void rtrs_from_imm(u32 imm, u32 *type, u32 *payload)
346 {
347 *payload = imm & MAX_IMM_PAYL_MASK;
348 *type = imm >> MAX_IMM_PAYL_BITS;
349 }
350
rtrs_to_io_req_imm(u32 addr)351 static inline u32 rtrs_to_io_req_imm(u32 addr)
352 {
353 return rtrs_to_imm(RTRS_IO_REQ_IMM, addr);
354 }
355
rtrs_to_io_rsp_imm(u32 msg_id,int errno,bool w_inval)356 static inline u32 rtrs_to_io_rsp_imm(u32 msg_id, int errno, bool w_inval)
357 {
358 enum rtrs_imm_type type;
359 u32 payload;
360
361 /* 9 bits for errno, 19 bits for msg_id */
362 payload = (abs(errno) & 0x1ff) << 19 | (msg_id & 0x7ffff);
363 type = w_inval ? RTRS_IO_RSP_W_INV_IMM : RTRS_IO_RSP_IMM;
364
365 return rtrs_to_imm(type, payload);
366 }
367
rtrs_from_io_rsp_imm(u32 payload,u32 * msg_id,int * errno)368 static inline void rtrs_from_io_rsp_imm(u32 payload, u32 *msg_id, int *errno)
369 {
370 /* 9 bits for errno, 19 bits for msg_id */
371 *msg_id = payload & 0x7ffff;
372 *errno = -(int)((payload >> 19) & 0x1ff);
373 }
374
375 #define STAT_STORE_FUNC(type, set_value, reset) \
376 static ssize_t set_value##_store(struct kobject *kobj, \
377 struct kobj_attribute *attr, \
378 const char *buf, size_t count) \
379 { \
380 int ret = -EINVAL; \
381 type *stats = container_of(kobj, type, kobj_stats); \
382 \
383 if (sysfs_streq(buf, "1")) \
384 ret = reset(stats, true); \
385 else if (sysfs_streq(buf, "0")) \
386 ret = reset(stats, false); \
387 if (ret) \
388 return ret; \
389 \
390 return count; \
391 }
392
393 #define STAT_SHOW_FUNC(type, get_value, print) \
394 static ssize_t get_value##_show(struct kobject *kobj, \
395 struct kobj_attribute *attr, \
396 char *page) \
397 { \
398 type *stats = container_of(kobj, type, kobj_stats); \
399 \
400 return print(stats, page); \
401 }
402
403 #define STAT_ATTR(type, stat, print, reset) \
404 STAT_STORE_FUNC(type, stat, reset) \
405 STAT_SHOW_FUNC(type, stat, print) \
406 static struct kobj_attribute stat##_attr = __ATTR_RW(stat)
407
408 #endif /* RTRS_PRI_H */
409