4 #include <rdma/ib_verbs.h>
5 #include <rdma/rdma_cm.h>
7 #include "rdma_transport.h"
9 #define RDS_FMR_SIZE 256
10 #define RDS_FMR_POOL_SIZE 4096
12 #define RDS_IB_MAX_SGE 8
13 #define RDS_IB_RECV_SGE 2
15 #define RDS_IB_DEFAULT_RECV_WR 1024
16 #define RDS_IB_DEFAULT_SEND_WR 256
18 #define RDS_IB_DEFAULT_RETRY_COUNT 2
20 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
22 extern struct list_head rds_ib_devices;
25 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
26 * try and minimize the amount of memory tied up both the device and
27 * socket receive queues.
29 /* page offset of the final full frag that fits in the page */
30 #define RDS_PAGE_LAST_OFF (((PAGE_SIZE / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE)
31 struct rds_page_frag {
32 struct list_head f_item;
34 unsigned long f_offset;
38 struct rds_ib_incoming {
39 struct list_head ii_frags;
40 struct rds_incoming ii_inc;
43 struct rds_ib_connect_private {
44 /* Add new fields at the end, and don't permute existing fields. */
49 __be16 dp_protocol_minor_mask; /* bitmask */
52 __be32 dp_credit; /* non-zero enables flow ctl */
55 struct rds_ib_send_work {
56 struct rds_message *s_rm;
57 struct rds_rdma_op *s_op;
58 struct ib_send_wr s_wr;
59 struct ib_sge s_sge[RDS_IB_MAX_SGE];
60 unsigned long s_queued;
63 struct rds_ib_recv_work {
64 struct rds_ib_incoming *r_ibinc;
65 struct rds_page_frag *r_frag;
66 struct ib_recv_wr r_wr;
67 struct ib_sge r_sge[2];
70 struct rds_ib_work_ring {
80 struct rds_ib_connection {
82 struct list_head ib_node;
83 struct rds_ib_device *rds_ibdev;
84 struct rds_connection *conn;
86 /* alphabet soup, IBTA style */
87 struct rdma_cm_id *i_cm_id;
90 struct ib_cq *i_send_cq;
91 struct ib_cq *i_recv_cq;
94 struct rds_ib_work_ring i_send_ring;
95 struct rds_message *i_rm;
96 struct rds_header *i_send_hdrs;
98 struct rds_ib_send_work *i_sends;
101 struct mutex i_recv_mutex;
102 struct rds_ib_work_ring i_recv_ring;
103 struct rds_ib_incoming *i_ibinc;
105 struct rds_header *i_recv_hdrs;
107 struct rds_ib_recv_work *i_recvs;
108 struct rds_page_frag i_frag;
109 u64 i_ack_recv; /* last ACK received */
112 unsigned long i_ack_flags;
113 #ifdef KERNEL_HAS_ATOMIC64
114 atomic64_t i_ack_next; /* next ACK to send */
116 spinlock_t i_ack_lock; /* protect i_ack_next */
117 u64 i_ack_next; /* next ACK to send */
119 struct rds_header *i_ack;
120 struct ib_send_wr i_ack_wr;
121 struct ib_sge i_ack_sge;
123 unsigned long i_ack_queued;
125 /* Flow control related information
127 * Our algorithm uses a pair variables that we need to access
128 * atomically - one for the send credits, and one posted
129 * recv credits we need to transfer to remote.
130 * Rather than protect them using a slow spinlock, we put both into
131 * a single atomic_t and update it using cmpxchg
135 /* Protocol version specific information */
136 unsigned int i_flowctl:1; /* enable/disable flow ctl */
138 /* Batched completions */
139 unsigned int i_unsignaled_wrs;
140 long i_unsignaled_bytes;
143 /* This assumes that atomic_t is at least 32 bits */
144 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
145 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
146 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
147 #define IB_SET_POST_CREDITS(v) ((v) << 16)
149 struct rds_ib_ipaddr {
150 struct list_head list;
154 struct rds_ib_device {
155 struct list_head list;
156 struct list_head ipaddr_list;
157 struct list_head conn_list;
158 struct ib_device *dev;
161 struct rds_ib_mr_pool *mr_pool;
165 unsigned int fmr_max_remaps;
166 unsigned int max_fmrs;
168 unsigned int max_wrs;
169 spinlock_t spinlock; /* protect the above */
172 /* bits for i_ack_flags */
173 #define IB_ACK_IN_FLIGHT 0
174 #define IB_ACK_REQUESTED 1
176 /* Magic WR_ID for ACKs */
177 #define RDS_IB_ACK_WR_ID (~(u64) 0)
179 struct rds_ib_statistics {
180 uint64_t s_ib_connect_raced;
181 uint64_t s_ib_listen_closed_stale;
182 uint64_t s_ib_tx_cq_call;
183 uint64_t s_ib_tx_cq_event;
184 uint64_t s_ib_tx_ring_full;
185 uint64_t s_ib_tx_throttle;
186 uint64_t s_ib_tx_sg_mapping_failure;
187 uint64_t s_ib_tx_stalled;
188 uint64_t s_ib_tx_credit_updates;
189 uint64_t s_ib_rx_cq_call;
190 uint64_t s_ib_rx_cq_event;
191 uint64_t s_ib_rx_ring_empty;
192 uint64_t s_ib_rx_refill_from_cq;
193 uint64_t s_ib_rx_refill_from_thread;
194 uint64_t s_ib_rx_alloc_limit;
195 uint64_t s_ib_rx_credit_updates;
196 uint64_t s_ib_ack_sent;
197 uint64_t s_ib_ack_send_failure;
198 uint64_t s_ib_ack_send_delayed;
199 uint64_t s_ib_ack_send_piggybacked;
200 uint64_t s_ib_ack_received;
201 uint64_t s_ib_rdma_mr_alloc;
202 uint64_t s_ib_rdma_mr_free;
203 uint64_t s_ib_rdma_mr_used;
204 uint64_t s_ib_rdma_mr_pool_flush;
205 uint64_t s_ib_rdma_mr_pool_wait;
206 uint64_t s_ib_rdma_mr_pool_depleted;
209 extern struct workqueue_struct *rds_ib_wq;
212 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
215 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
216 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
220 for (i = 0; i < sg_dma_len; ++i) {
221 ib_dma_sync_single_for_cpu(dev,
222 ib_sg_dma_address(dev, &sg[i]),
223 ib_sg_dma_len(dev, &sg[i]),
227 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
229 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
230 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
234 for (i = 0; i < sg_dma_len; ++i) {
235 ib_dma_sync_single_for_device(dev,
236 ib_sg_dma_address(dev, &sg[i]),
237 ib_sg_dma_len(dev, &sg[i]),
241 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
245 extern struct rds_transport rds_ib_transport;
246 extern void rds_ib_add_one(struct ib_device *device);
247 extern void rds_ib_remove_one(struct ib_device *device);
248 extern struct ib_client rds_ib_client;
250 extern unsigned int fmr_pool_size;
251 extern unsigned int fmr_message_size;
252 extern unsigned int rds_ib_retry_count;
254 extern spinlock_t ib_nodev_conns_lock;
255 extern struct list_head ib_nodev_conns;
258 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
259 void rds_ib_conn_free(void *arg);
260 int rds_ib_conn_connect(struct rds_connection *conn);
261 void rds_ib_conn_shutdown(struct rds_connection *conn);
262 void rds_ib_state_change(struct sock *sk);
263 int __init rds_ib_listen_init(void);
264 void rds_ib_listen_stop(void);
265 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
266 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
267 struct rdma_cm_event *event);
268 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
269 void rds_ib_cm_connect_complete(struct rds_connection *conn,
270 struct rdma_cm_event *event);
273 #define rds_ib_conn_error(conn, fmt...) \
274 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
277 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
278 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
279 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
280 void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock);
281 static inline void rds_ib_destroy_nodev_conns(void)
283 __rds_ib_destroy_conns(&ib_nodev_conns, &ib_nodev_conns_lock);
285 static inline void rds_ib_destroy_conns(struct rds_ib_device *rds_ibdev)
287 __rds_ib_destroy_conns(&rds_ibdev->conn_list, &rds_ibdev->spinlock);
289 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
290 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
291 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
292 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
293 struct rds_sock *rs, u32 *key_ret);
294 void rds_ib_sync_mr(void *trans_private, int dir);
295 void rds_ib_free_mr(void *trans_private, int invalidate);
296 void rds_ib_flush_mrs(void);
299 int __init rds_ib_recv_init(void);
300 void rds_ib_recv_exit(void);
301 int rds_ib_recv(struct rds_connection *conn);
302 int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
303 gfp_t page_gfp, int prefill);
304 void rds_ib_inc_purge(struct rds_incoming *inc);
305 void rds_ib_inc_free(struct rds_incoming *inc);
306 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
308 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
309 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
310 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
311 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
312 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
313 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
314 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
317 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
318 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
319 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
320 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
321 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
322 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
323 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
324 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
325 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
326 extern wait_queue_head_t rds_ib_ring_empty_wait;
329 void rds_ib_xmit_complete(struct rds_connection *conn);
330 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
331 unsigned int hdr_off, unsigned int sg, unsigned int off);
332 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
333 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
334 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
335 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op);
336 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
337 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
338 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
339 u32 *adv_credits, int need_posted, int max_posted);
342 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
343 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
344 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
348 int __init rds_ib_sysctl_init(void);
349 void rds_ib_sysctl_exit(void);
350 extern unsigned long rds_ib_sysctl_max_send_wr;
351 extern unsigned long rds_ib_sysctl_max_recv_wr;
352 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
353 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
354 extern unsigned long rds_ib_sysctl_max_recv_allocation;
355 extern unsigned int rds_ib_sysctl_flow_control;
356 extern ctl_table rds_ib_sysctl_table[];
359 * Helper functions for getting/setting the header and data SGEs in
360 * RDS packets (not RDMA)
362 * From version 3.1 onwards, header is in front of data in the sge.
364 static inline struct ib_sge *
365 rds_ib_header_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
367 if (ic->conn->c_version > RDS_PROTOCOL_3_0)
373 static inline struct ib_sge *
374 rds_ib_data_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
376 if (ic->conn->c_version > RDS_PROTOCOL_3_0)