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 u16 cq_size; 95 }; 96 97 struct rtrs_sess { 98 struct list_head entry; 99 struct sockaddr_storage dst_addr; 100 struct sockaddr_storage src_addr; 101 char sessname[NAME_MAX]; 102 uuid_t uuid; 103 struct rtrs_con **con; 104 unsigned int con_num; 105 unsigned int irq_con_num; 106 unsigned int recon_cnt; 107 struct rtrs_ib_dev *dev; 108 int dev_ref; 109 struct ib_cqe *hb_cqe; 110 void (*hb_err_handler)(struct rtrs_con *con); 111 struct workqueue_struct *hb_wq; 112 struct delayed_work hb_dwork; 113 unsigned int hb_interval_ms; 114 unsigned int hb_missed_cnt; 115 unsigned int hb_missed_max; 116 ktime_t hb_last_sent; 117 ktime_t hb_cur_latency; 118 }; 119 120 /* rtrs information unit */ 121 struct rtrs_iu { 122 struct ib_cqe cqe; 123 dma_addr_t dma_addr; 124 void *buf; 125 size_t size; 126 enum dma_data_direction direction; 127 }; 128 129 /** 130 * enum rtrs_msg_types - RTRS message types, see also rtrs/README 131 * @RTRS_MSG_INFO_REQ: Client additional info request to the server 132 * @RTRS_MSG_INFO_RSP: Server additional info response to the client 133 * @RTRS_MSG_WRITE: Client writes data per RDMA to server 134 * @RTRS_MSG_READ: Client requests data transfer from server 135 * @RTRS_MSG_RKEY_RSP: Server refreshed rkey for rbuf 136 */ 137 enum rtrs_msg_types { 138 RTRS_MSG_INFO_REQ, 139 RTRS_MSG_INFO_RSP, 140 RTRS_MSG_WRITE, 141 RTRS_MSG_READ, 142 RTRS_MSG_RKEY_RSP, 143 }; 144 145 /** 146 * enum rtrs_msg_flags - RTRS message flags. 147 * @RTRS_NEED_INVAL: Send invalidation in response. 148 * @RTRS_MSG_NEW_RKEY_F: Send refreshed rkey in response. 149 */ 150 enum rtrs_msg_flags { 151 RTRS_MSG_NEED_INVAL_F = 1 << 0, 152 RTRS_MSG_NEW_RKEY_F = 1 << 1, 153 }; 154 155 /** 156 * struct rtrs_sg_desc - RDMA-Buffer entry description 157 * @addr: Address of RDMA destination buffer 158 * @key: Authorization rkey to write to the buffer 159 * @len: Size of the buffer 160 */ 161 struct rtrs_sg_desc { 162 __le64 addr; 163 __le32 key; 164 __le32 len; 165 }; 166 167 /** 168 * struct rtrs_msg_conn_req - Client connection request to the server 169 * @magic: RTRS magic 170 * @version: RTRS protocol version 171 * @cid: Current connection id 172 * @cid_num: Number of connections per session 173 * @recon_cnt: Reconnections counter 174 * @sess_uuid: UUID of a session (path) 175 * @paths_uuid: UUID of a group of sessions (paths) 176 * 177 * NOTE: max size 56 bytes, see man rdma_connect(). 178 */ 179 struct rtrs_msg_conn_req { 180 /* Is set to 0 by cma.c in case of AF_IB, do not touch that. 181 * see https://www.spinics.net/lists/linux-rdma/msg22397.html 182 */ 183 u8 __cma_version; 184 /* On sender side that should be set to 0, or cma_save_ip_info() 185 * extract garbage and will fail. 186 */ 187 u8 __ip_version; 188 __le16 magic; 189 __le16 version; 190 __le16 cid; 191 __le16 cid_num; 192 __le16 recon_cnt; 193 uuid_t sess_uuid; 194 uuid_t paths_uuid; 195 u8 first_conn : 1; 196 u8 reserved_bits : 7; 197 u8 reserved[11]; 198 }; 199 200 /** 201 * struct rtrs_msg_conn_rsp - Server connection response to the client 202 * @magic: RTRS magic 203 * @version: RTRS protocol version 204 * @errno: If rdma_accept() then 0, if rdma_reject() indicates error 205 * @queue_depth: max inflight messages (queue-depth) in this session 206 * @max_io_size: max io size server supports 207 * @max_hdr_size: max msg header size server supports 208 * 209 * NOTE: size is 56 bytes, max possible is 136 bytes, see man rdma_accept(). 210 */ 211 struct rtrs_msg_conn_rsp { 212 __le16 magic; 213 __le16 version; 214 __le16 errno; 215 __le16 queue_depth; 216 __le32 max_io_size; 217 __le32 max_hdr_size; 218 __le32 flags; 219 u8 reserved[36]; 220 }; 221 222 /** 223 * struct rtrs_msg_info_req 224 * @type: @RTRS_MSG_INFO_REQ 225 * @sessname: Session name chosen by client 226 */ 227 struct rtrs_msg_info_req { 228 __le16 type; 229 u8 sessname[NAME_MAX]; 230 u8 reserved[15]; 231 }; 232 233 /** 234 * struct rtrs_msg_info_rsp 235 * @type: @RTRS_MSG_INFO_RSP 236 * @sg_cnt: Number of @desc entries 237 * @desc: RDMA buffers where the client can write to server 238 */ 239 struct rtrs_msg_info_rsp { 240 __le16 type; 241 __le16 sg_cnt; 242 u8 reserved[4]; 243 struct rtrs_sg_desc desc[]; 244 }; 245 246 /** 247 * struct rtrs_msg_rkey_rsp 248 * @type: @RTRS_MSG_RKEY_RSP 249 * @buf_id: RDMA buf_id of the new rkey 250 * @rkey: new remote key for RDMA buffers id from server 251 */ 252 struct rtrs_msg_rkey_rsp { 253 __le16 type; 254 __le16 buf_id; 255 __le32 rkey; 256 }; 257 258 /** 259 * struct rtrs_msg_rdma_read - RDMA data transfer request from client 260 * @type: always @RTRS_MSG_READ 261 * @usr_len: length of user payload 262 * @sg_cnt: number of @desc entries 263 * @desc: RDMA buffers where the server can write the result to 264 */ 265 struct rtrs_msg_rdma_read { 266 __le16 type; 267 __le16 usr_len; 268 __le16 flags; 269 __le16 sg_cnt; 270 struct rtrs_sg_desc desc[]; 271 }; 272 273 /** 274 * struct_msg_rdma_write - Message transferred to server with RDMA-Write 275 * @type: always @RTRS_MSG_WRITE 276 * @usr_len: length of user payload 277 */ 278 struct rtrs_msg_rdma_write { 279 __le16 type; 280 __le16 usr_len; 281 }; 282 283 /** 284 * struct_msg_rdma_hdr - header for read or write request 285 * @type: @RTRS_MSG_WRITE | @RTRS_MSG_READ 286 */ 287 struct rtrs_msg_rdma_hdr { 288 __le16 type; 289 }; 290 291 /* rtrs.c */ 292 293 struct rtrs_iu *rtrs_iu_alloc(u32 queue_size, size_t size, gfp_t t, 294 struct ib_device *dev, enum dma_data_direction, 295 void (*done)(struct ib_cq *cq, struct ib_wc *wc)); 296 void rtrs_iu_free(struct rtrs_iu *iu, struct ib_device *dev, u32 queue_size); 297 int rtrs_iu_post_recv(struct rtrs_con *con, struct rtrs_iu *iu); 298 int rtrs_iu_post_send(struct rtrs_con *con, struct rtrs_iu *iu, size_t size, 299 struct ib_send_wr *head); 300 int rtrs_iu_post_rdma_write_imm(struct rtrs_con *con, struct rtrs_iu *iu, 301 struct ib_sge *sge, unsigned int num_sge, 302 u32 rkey, u64 rdma_addr, u32 imm_data, 303 enum ib_send_flags flags, 304 struct ib_send_wr *head); 305 306 int rtrs_post_recv_empty(struct rtrs_con *con, struct ib_cqe *cqe); 307 int rtrs_post_rdma_write_imm_empty(struct rtrs_con *con, struct ib_cqe *cqe, 308 u32 imm_data, enum ib_send_flags flags, 309 struct ib_send_wr *head); 310 311 int rtrs_cq_qp_create(struct rtrs_sess *rtrs_sess, struct rtrs_con *con, 312 u32 max_send_sge, int cq_vector, int cq_size, 313 u32 max_send_wr, u32 max_recv_wr, 314 enum ib_poll_context poll_ctx); 315 void rtrs_cq_qp_destroy(struct rtrs_con *con); 316 317 void rtrs_init_hb(struct rtrs_sess *sess, struct ib_cqe *cqe, 318 unsigned int interval_ms, unsigned int missed_max, 319 void (*err_handler)(struct rtrs_con *con), 320 struct workqueue_struct *wq); 321 void rtrs_start_hb(struct rtrs_sess *sess); 322 void rtrs_stop_hb(struct rtrs_sess *sess); 323 void rtrs_send_hb_ack(struct rtrs_sess *sess); 324 325 void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags, 326 struct rtrs_rdma_dev_pd *pool); 327 void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool); 328 329 struct rtrs_ib_dev *rtrs_ib_dev_find_or_add(struct ib_device *ib_dev, 330 struct rtrs_rdma_dev_pd *pool); 331 int rtrs_ib_dev_put(struct rtrs_ib_dev *dev); 332 333 static inline u32 rtrs_to_imm(u32 type, u32 payload) 334 { 335 BUILD_BUG_ON(MAX_IMM_PAYL_BITS + MAX_IMM_TYPE_BITS != 32); 336 BUILD_BUG_ON(RTRS_LAST_IMM > (1<<MAX_IMM_TYPE_BITS)); 337 return ((type & MAX_IMM_TYPE_MASK) << MAX_IMM_PAYL_BITS) | 338 (payload & MAX_IMM_PAYL_MASK); 339 } 340 341 static inline void rtrs_from_imm(u32 imm, u32 *type, u32 *payload) 342 { 343 *payload = imm & MAX_IMM_PAYL_MASK; 344 *type = imm >> MAX_IMM_PAYL_BITS; 345 } 346 347 static inline u32 rtrs_to_io_req_imm(u32 addr) 348 { 349 return rtrs_to_imm(RTRS_IO_REQ_IMM, addr); 350 } 351 352 static inline u32 rtrs_to_io_rsp_imm(u32 msg_id, int errno, bool w_inval) 353 { 354 enum rtrs_imm_type type; 355 u32 payload; 356 357 /* 9 bits for errno, 19 bits for msg_id */ 358 payload = (abs(errno) & 0x1ff) << 19 | (msg_id & 0x7ffff); 359 type = w_inval ? RTRS_IO_RSP_W_INV_IMM : RTRS_IO_RSP_IMM; 360 361 return rtrs_to_imm(type, payload); 362 } 363 364 static inline void rtrs_from_io_rsp_imm(u32 payload, u32 *msg_id, int *errno) 365 { 366 /* 9 bits for errno, 19 bits for msg_id */ 367 *msg_id = payload & 0x7ffff; 368 *errno = -(int)((payload >> 19) & 0x1ff); 369 } 370 371 #define STAT_STORE_FUNC(type, set_value, reset) \ 372 static ssize_t set_value##_store(struct kobject *kobj, \ 373 struct kobj_attribute *attr, \ 374 const char *buf, size_t count) \ 375 { \ 376 int ret = -EINVAL; \ 377 type *stats = container_of(kobj, type, kobj_stats); \ 378 \ 379 if (sysfs_streq(buf, "1")) \ 380 ret = reset(stats, true); \ 381 else if (sysfs_streq(buf, "0")) \ 382 ret = reset(stats, false); \ 383 if (ret) \ 384 return ret; \ 385 \ 386 return count; \ 387 } 388 389 #define STAT_SHOW_FUNC(type, get_value, print) \ 390 static ssize_t get_value##_show(struct kobject *kobj, \ 391 struct kobj_attribute *attr, \ 392 char *page) \ 393 { \ 394 type *stats = container_of(kobj, type, kobj_stats); \ 395 \ 396 return print(stats, page, PAGE_SIZE); \ 397 } 398 399 #define STAT_ATTR(type, stat, print, reset) \ 400 STAT_STORE_FUNC(type, stat, reset) \ 401 STAT_SHOW_FUNC(type, stat, print) \ 402 static struct kobj_attribute stat##_attr = __ATTR_RW(stat) 403 404 #endif /* RTRS_PRI_H */ 405