4 #include <rdma/ib_verbs.h>
5 #include <rdma/rdma_cm.h>
7 #include <linux/slab.h>
9 #include "rdma_transport.h"
11 #define RDS_FMR_SIZE 256
12 #define RDS_FMR_POOL_SIZE 8192
14 #define RDS_IB_MAX_SGE 8
15 #define RDS_IB_RECV_SGE 2
17 #define RDS_IB_DEFAULT_RECV_WR 1024
18 #define RDS_IB_DEFAULT_SEND_WR 256
20 #define RDS_IB_DEFAULT_RETRY_COUNT 2
22 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
24 #define RDS_IB_RECYCLE_BATCH_COUNT 32
26 extern struct list_head rds_ib_devices;
29 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
30 * try and minimize the amount of memory tied up both the device and
31 * socket receive queues.
33 struct rds_page_frag {
34 struct list_head f_item;
35 struct list_head f_cache_entry;
36 struct scatterlist f_sg;
39 struct rds_ib_incoming {
40 struct list_head ii_frags;
41 struct list_head ii_cache_entry;
42 struct rds_incoming ii_inc;
45 struct rds_ib_cache_head {
46 struct list_head *first;
50 struct rds_ib_refill_cache {
51 struct rds_ib_cache_head *percpu;
52 struct list_head *xfer;
53 struct list_head *ready;
56 struct rds_ib_connect_private {
57 /* Add new fields at the end, and don't permute existing fields. */
62 __be16 dp_protocol_minor_mask; /* bitmask */
65 __be32 dp_credit; /* non-zero enables flow ctl */
68 struct rds_ib_send_work {
70 struct ib_send_wr s_wr;
71 struct ib_sge s_sge[RDS_IB_MAX_SGE];
72 unsigned long s_queued;
75 struct rds_ib_recv_work {
76 struct rds_ib_incoming *r_ibinc;
77 struct rds_page_frag *r_frag;
78 struct ib_recv_wr r_wr;
79 struct ib_sge r_sge[2];
82 struct rds_ib_work_ring {
92 struct rds_ib_connection {
94 struct list_head ib_node;
95 struct rds_ib_device *rds_ibdev;
96 struct rds_connection *conn;
98 /* alphabet soup, IBTA style */
99 struct rdma_cm_id *i_cm_id;
102 struct ib_cq *i_send_cq;
103 struct ib_cq *i_recv_cq;
106 struct rds_ib_work_ring i_send_ring;
107 struct rm_data_op *i_data_op;
108 struct rds_header *i_send_hdrs;
110 struct rds_ib_send_work *i_sends;
111 atomic_t i_signaled_sends;
114 struct tasklet_struct i_recv_tasklet;
115 struct mutex i_recv_mutex;
116 struct rds_ib_work_ring i_recv_ring;
117 struct rds_ib_incoming *i_ibinc;
119 struct rds_header *i_recv_hdrs;
121 struct rds_ib_recv_work *i_recvs;
122 u64 i_ack_recv; /* last ACK received */
123 struct rds_ib_refill_cache i_cache_incs;
124 struct rds_ib_refill_cache i_cache_frags;
127 unsigned long i_ack_flags;
128 #ifdef KERNEL_HAS_ATOMIC64
129 atomic64_t i_ack_next; /* next ACK to send */
131 spinlock_t i_ack_lock; /* protect i_ack_next */
132 u64 i_ack_next; /* next ACK to send */
134 struct rds_header *i_ack;
135 struct ib_send_wr i_ack_wr;
136 struct ib_sge i_ack_sge;
138 unsigned long i_ack_queued;
140 /* Flow control related information
142 * Our algorithm uses a pair variables that we need to access
143 * atomically - one for the send credits, and one posted
144 * recv credits we need to transfer to remote.
145 * Rather than protect them using a slow spinlock, we put both into
146 * a single atomic_t and update it using cmpxchg
150 /* Protocol version specific information */
151 unsigned int i_flowctl:1; /* enable/disable flow ctl */
153 /* Batched completions */
154 unsigned int i_unsignaled_wrs;
157 /* This assumes that atomic_t is at least 32 bits */
158 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
159 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
160 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
161 #define IB_SET_POST_CREDITS(v) ((v) << 16)
163 struct rds_ib_ipaddr {
164 struct list_head list;
168 struct rds_ib_device {
169 struct list_head list;
170 struct list_head ipaddr_list;
171 struct list_head conn_list;
172 struct ib_device *dev;
175 struct rds_ib_mr_pool *mr_pool;
176 unsigned int fmr_max_remaps;
177 unsigned int max_fmrs;
179 unsigned int max_wrs;
180 unsigned int max_initiator_depth;
181 unsigned int max_responder_resources;
182 spinlock_t spinlock; /* protect the above */
184 struct work_struct free_work;
187 #define pcidev_to_node(pcidev) pcibus_to_node(pcidev->bus)
188 #define ibdev_to_node(ibdev) pcidev_to_node(to_pci_dev(ibdev->dma_device))
189 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
191 /* bits for i_ack_flags */
192 #define IB_ACK_IN_FLIGHT 0
193 #define IB_ACK_REQUESTED 1
195 /* Magic WR_ID for ACKs */
196 #define RDS_IB_ACK_WR_ID (~(u64) 0)
198 struct rds_ib_statistics {
199 uint64_t s_ib_connect_raced;
200 uint64_t s_ib_listen_closed_stale;
201 uint64_t s_ib_tx_cq_call;
202 uint64_t s_ib_tx_cq_event;
203 uint64_t s_ib_tx_ring_full;
204 uint64_t s_ib_tx_throttle;
205 uint64_t s_ib_tx_sg_mapping_failure;
206 uint64_t s_ib_tx_stalled;
207 uint64_t s_ib_tx_credit_updates;
208 uint64_t s_ib_rx_cq_call;
209 uint64_t s_ib_rx_cq_event;
210 uint64_t s_ib_rx_ring_empty;
211 uint64_t s_ib_rx_refill_from_cq;
212 uint64_t s_ib_rx_refill_from_thread;
213 uint64_t s_ib_rx_alloc_limit;
214 uint64_t s_ib_rx_credit_updates;
215 uint64_t s_ib_ack_sent;
216 uint64_t s_ib_ack_send_failure;
217 uint64_t s_ib_ack_send_delayed;
218 uint64_t s_ib_ack_send_piggybacked;
219 uint64_t s_ib_ack_received;
220 uint64_t s_ib_rdma_mr_alloc;
221 uint64_t s_ib_rdma_mr_free;
222 uint64_t s_ib_rdma_mr_used;
223 uint64_t s_ib_rdma_mr_pool_flush;
224 uint64_t s_ib_rdma_mr_pool_wait;
225 uint64_t s_ib_rdma_mr_pool_depleted;
226 uint64_t s_ib_atomic_cswp;
227 uint64_t s_ib_atomic_fadd;
230 extern struct workqueue_struct *rds_ib_wq;
233 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
236 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
237 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
241 for (i = 0; i < sg_dma_len; ++i) {
242 ib_dma_sync_single_for_cpu(dev,
243 ib_sg_dma_address(dev, &sg[i]),
244 ib_sg_dma_len(dev, &sg[i]),
248 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
250 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
251 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
255 for (i = 0; i < sg_dma_len; ++i) {
256 ib_dma_sync_single_for_device(dev,
257 ib_sg_dma_address(dev, &sg[i]),
258 ib_sg_dma_len(dev, &sg[i]),
262 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
266 extern struct rds_transport rds_ib_transport;
267 extern void rds_ib_add_one(struct ib_device *device);
268 extern void rds_ib_remove_one(struct ib_device *device);
269 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
270 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
271 extern struct ib_client rds_ib_client;
273 extern unsigned int fmr_pool_size;
274 extern unsigned int fmr_message_size;
275 extern unsigned int rds_ib_retry_count;
277 extern spinlock_t ib_nodev_conns_lock;
278 extern struct list_head ib_nodev_conns;
281 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
282 void rds_ib_conn_free(void *arg);
283 int rds_ib_conn_connect(struct rds_connection *conn);
284 void rds_ib_conn_shutdown(struct rds_connection *conn);
285 void rds_ib_state_change(struct sock *sk);
286 int rds_ib_listen_init(void);
287 void rds_ib_listen_stop(void);
288 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
289 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
290 struct rdma_cm_event *event);
291 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
292 void rds_ib_cm_connect_complete(struct rds_connection *conn,
293 struct rdma_cm_event *event);
296 #define rds_ib_conn_error(conn, fmt...) \
297 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
300 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
301 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
302 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
303 void rds_ib_destroy_nodev_conns(void);
304 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
305 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
306 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
307 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
308 struct rds_sock *rs, u32 *key_ret);
309 void rds_ib_sync_mr(void *trans_private, int dir);
310 void rds_ib_free_mr(void *trans_private, int invalidate);
311 void rds_ib_flush_mrs(void);
312 int rds_ib_fmr_init(void);
313 void rds_ib_fmr_exit(void);
316 int rds_ib_recv_init(void);
317 void rds_ib_recv_exit(void);
318 int rds_ib_recv(struct rds_connection *conn);
319 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic);
320 void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
321 void rds_ib_recv_refill(struct rds_connection *conn, int prefill);
322 void rds_ib_inc_free(struct rds_incoming *inc);
323 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
325 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
326 void rds_ib_recv_tasklet_fn(unsigned long data);
327 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
328 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
329 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
330 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
331 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
332 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
335 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
336 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
337 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
338 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
339 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
340 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
341 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
342 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
343 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
344 extern wait_queue_head_t rds_ib_ring_empty_wait;
347 void rds_ib_xmit_complete(struct rds_connection *conn);
348 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
349 unsigned int hdr_off, unsigned int sg, unsigned int off);
350 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
351 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
352 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
353 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
354 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
355 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
356 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
357 u32 *adv_credits, int need_posted, int max_posted);
358 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
361 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
362 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
363 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
367 int rds_ib_sysctl_init(void);
368 void rds_ib_sysctl_exit(void);
369 extern unsigned long rds_ib_sysctl_max_send_wr;
370 extern unsigned long rds_ib_sysctl_max_recv_wr;
371 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
372 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
373 extern unsigned long rds_ib_sysctl_max_recv_allocation;
374 extern unsigned int rds_ib_sysctl_flow_control;
375 extern ctl_table rds_ib_sysctl_table[];