1 /* SPDX-License-Identifier: GPL-2.0-or-later */
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.
13 #include <linux/uuid.h>
14 #include <rdma/rdma_cm.h>
15 #include <rdma/ib_verbs.h>
20 #define RTRS_PROTO_VER_MAJOR 2
21 #define RTRS_PROTO_VER_MINOR 0
23 #define RTRS_PROTO_VER_STRING __stringify(RTRS_PROTO_VER_MAJOR) "." \
24 __stringify(RTRS_PROTO_VER_MINOR)
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.
35 #define MAX_SESS_QUEUE_DEPTH 65535
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),
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 */
49 RTRS_HB_MSG_IMM = 8, /* HB: HeartBeat */
56 SERVICE_CON_QUEUE_DEPTH = 512,
60 MIN_CHUNK_SIZE = 8192,
62 RTRS_HB_INTERVAL_MS = 5000,
63 RTRS_HB_MISSED_MAX = 5,
66 RTRS_PROTO_VER = (RTRS_PROTO_VER_MAJOR << 8) | RTRS_PROTO_VER_MINOR,
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);
78 struct rtrs_rdma_dev_pd {
80 struct list_head list;
81 enum ib_pd_flags pd_flags;
82 const struct rtrs_rdma_dev_pd_ops *ops;
86 struct ib_device *ib_dev;
89 struct list_head entry;
90 struct rtrs_rdma_dev_pd *pool;
94 struct rtrs_sess *sess;
97 struct rdma_cm_id *cm_id;
102 struct list_head entry;
103 struct sockaddr_storage dst_addr;
104 struct sockaddr_storage src_addr;
105 char sessname[NAME_MAX];
107 struct rtrs_con **con;
108 unsigned int con_num;
109 unsigned int recon_cnt;
110 struct rtrs_ib_dev *dev;
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;
121 /* rtrs information unit */
127 enum dma_data_direction direction;
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
138 enum rtrs_msg_types {
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.
151 enum rtrs_msg_flags {
152 RTRS_MSG_NEED_INVAL_F = 1 << 0,
153 RTRS_MSG_NEW_RKEY_F = 1 << 1,
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
162 struct rtrs_sg_desc {
169 * struct rtrs_msg_conn_req - Client connection request to the server
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)
178 * NOTE: max size 56 bytes, see man rdma_connect().
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
185 /* On sender side that should be set to 0, or cma_save_ip_info()
186 * extract garbage and will fail.
197 u8 reserved_bits : 7;
202 * struct rtrs_msg_conn_rsp - Server connection response to the client
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
210 * NOTE: size is 56 bytes, max possible is 136 bytes, see man rdma_accept().
212 struct rtrs_msg_conn_rsp {
224 * struct rtrs_msg_info_req
225 * @type: @RTRS_MSG_INFO_REQ
226 * @sessname: Session name chosen by client
228 struct rtrs_msg_info_req {
230 u8 sessname[NAME_MAX];
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
240 struct rtrs_msg_info_rsp {
244 struct rtrs_sg_desc desc[];
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
253 struct rtrs_msg_rkey_rsp {
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
266 struct rtrs_msg_rdma_read {
271 struct rtrs_sg_desc desc[];
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
279 struct rtrs_msg_rdma_write {
285 * struct_msg_rdma_hdr - header for read or write request
286 * @type: @RTRS_MSG_WRITE | @RTRS_MSG_READ
288 struct rtrs_msg_rdma_hdr {
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);
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);
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);
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);
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);
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);
334 static inline u32 rtrs_to_imm(u32 type, u32 payload)
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);
342 static inline void rtrs_from_imm(u32 imm, u32 *type, u32 *payload)
344 *payload = imm & MAX_IMM_PAYL_MASK;
345 *type = imm >> MAX_IMM_PAYL_BITS;
348 static inline u32 rtrs_to_io_req_imm(u32 addr)
350 return rtrs_to_imm(RTRS_IO_REQ_IMM, addr);
353 static inline u32 rtrs_to_io_rsp_imm(u32 msg_id, int errno, bool w_inval)
355 enum rtrs_imm_type type;
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;
362 return rtrs_to_imm(type, payload);
365 static inline void rtrs_from_io_rsp_imm(u32 payload, u32 *msg_id, int *errno)
367 /* 9 bits for errno, 19 bits for msg_id */
368 *msg_id = payload & 0x7ffff;
369 *errno = -(int)((payload >> 19) & 0x1ff);
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) \
378 type *stats = container_of(kobj, type, kobj_stats); \
380 if (sysfs_streq(buf, "1")) \
381 ret = reset(stats, true); \
382 else if (sysfs_streq(buf, "0")) \
383 ret = reset(stats, false); \
390 #define STAT_SHOW_FUNC(type, get_value, print) \
391 static ssize_t get_value##_show(struct kobject *kobj, \
392 struct kobj_attribute *attr, \
395 type *stats = container_of(kobj, type, kobj_stats); \
397 return print(stats, page, PAGE_SIZE); \
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)
405 #endif /* RTRS_PRI_H */