1 #ifndef _RDS_IB_H
2 #define _RDS_IB_H
3
4 #include <rdma/ib_verbs.h>
5 #include <rdma/rdma_cm.h>
6 #include <linux/interrupt.h>
7 #include <linux/pci.h>
8 #include <linux/slab.h>
9 #include "rds.h"
10 #include "rdma_transport.h"
11
12 #define RDS_FMR_SIZE 256
13 #define RDS_FMR_POOL_SIZE 8192
14
15 #define RDS_IB_MAX_SGE 8
16 #define RDS_IB_RECV_SGE 2
17
18 #define RDS_IB_DEFAULT_RECV_WR 1024
19 #define RDS_IB_DEFAULT_SEND_WR 256
20
21 #define RDS_IB_DEFAULT_RETRY_COUNT 2
22
23 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
24
25 #define RDS_IB_RECYCLE_BATCH_COUNT 32
26
27 extern struct rw_semaphore rds_ib_devices_lock;
28 extern struct list_head rds_ib_devices;
29
30 /*
31 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
32 * try and minimize the amount of memory tied up both the device and
33 * socket receive queues.
34 */
35 struct rds_page_frag {
36 struct list_head f_item;
37 struct list_head f_cache_entry;
38 struct scatterlist f_sg;
39 };
40
41 struct rds_ib_incoming {
42 struct list_head ii_frags;
43 struct list_head ii_cache_entry;
44 struct rds_incoming ii_inc;
45 };
46
47 struct rds_ib_cache_head {
48 struct list_head *first;
49 unsigned long count;
50 };
51
52 struct rds_ib_refill_cache {
53 struct rds_ib_cache_head *percpu;
54 struct list_head *xfer;
55 struct list_head *ready;
56 };
57
58 struct rds_ib_connect_private {
59 /* Add new fields at the end, and don't permute existing fields. */
60 __be32 dp_saddr;
61 __be32 dp_daddr;
62 u8 dp_protocol_major;
63 u8 dp_protocol_minor;
64 __be16 dp_protocol_minor_mask; /* bitmask */
65 __be32 dp_reserved1;
66 __be64 dp_ack_seq;
67 __be32 dp_credit; /* non-zero enables flow ctl */
68 };
69
70 struct rds_ib_send_work {
71 void *s_op;
72 struct ib_send_wr s_wr;
73 struct ib_sge s_sge[RDS_IB_MAX_SGE];
74 unsigned long s_queued;
75 };
76
77 struct rds_ib_recv_work {
78 struct rds_ib_incoming *r_ibinc;
79 struct rds_page_frag *r_frag;
80 struct ib_recv_wr r_wr;
81 struct ib_sge r_sge[2];
82 };
83
84 struct rds_ib_work_ring {
85 u32 w_nr;
86 u32 w_alloc_ptr;
87 u32 w_alloc_ctr;
88 u32 w_free_ptr;
89 atomic_t w_free_ctr;
90 };
91
92 struct rds_ib_device;
93
94 struct rds_ib_connection {
95
96 struct list_head ib_node;
97 struct rds_ib_device *rds_ibdev;
98 struct rds_connection *conn;
99
100 /* alphabet soup, IBTA style */
101 struct rdma_cm_id *i_cm_id;
102 struct ib_pd *i_pd;
103 struct ib_mr *i_mr;
104 struct ib_cq *i_send_cq;
105 struct ib_cq *i_recv_cq;
106
107 /* tx */
108 struct rds_ib_work_ring i_send_ring;
109 struct rm_data_op *i_data_op;
110 struct rds_header *i_send_hdrs;
111 u64 i_send_hdrs_dma;
112 struct rds_ib_send_work *i_sends;
113 atomic_t i_signaled_sends;
114
115 /* rx */
116 struct tasklet_struct i_recv_tasklet;
117 struct mutex i_recv_mutex;
118 struct rds_ib_work_ring i_recv_ring;
119 struct rds_ib_incoming *i_ibinc;
120 u32 i_recv_data_rem;
121 struct rds_header *i_recv_hdrs;
122 u64 i_recv_hdrs_dma;
123 struct rds_ib_recv_work *i_recvs;
124 u64 i_ack_recv; /* last ACK received */
125 struct rds_ib_refill_cache i_cache_incs;
126 struct rds_ib_refill_cache i_cache_frags;
127
128 /* sending acks */
129 unsigned long i_ack_flags;
130 #ifdef KERNEL_HAS_ATOMIC64
131 atomic64_t i_ack_next; /* next ACK to send */
132 #else
133 spinlock_t i_ack_lock; /* protect i_ack_next */
134 u64 i_ack_next; /* next ACK to send */
135 #endif
136 struct rds_header *i_ack;
137 struct ib_send_wr i_ack_wr;
138 struct ib_sge i_ack_sge;
139 u64 i_ack_dma;
140 unsigned long i_ack_queued;
141
142 /* Flow control related information
143 *
144 * Our algorithm uses a pair variables that we need to access
145 * atomically - one for the send credits, and one posted
146 * recv credits we need to transfer to remote.
147 * Rather than protect them using a slow spinlock, we put both into
148 * a single atomic_t and update it using cmpxchg
149 */
150 atomic_t i_credits;
151
152 /* Protocol version specific information */
153 unsigned int i_flowctl:1; /* enable/disable flow ctl */
154
155 /* Batched completions */
156 unsigned int i_unsignaled_wrs;
157 };
158
159 /* This assumes that atomic_t is at least 32 bits */
160 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
161 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
162 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
163 #define IB_SET_POST_CREDITS(v) ((v) << 16)
164
165 struct rds_ib_ipaddr {
166 struct list_head list;
167 __be32 ipaddr;
168 };
169
170 struct rds_ib_device {
171 struct list_head list;
172 struct list_head ipaddr_list;
173 struct list_head conn_list;
174 struct ib_device *dev;
175 struct ib_pd *pd;
176 struct ib_mr *mr;
177 struct rds_ib_mr_pool *mr_pool;
178 unsigned int fmr_max_remaps;
179 unsigned int max_fmrs;
180 int max_sge;
181 unsigned int max_wrs;
182 unsigned int max_initiator_depth;
183 unsigned int max_responder_resources;
184 spinlock_t spinlock; /* protect the above */
185 atomic_t refcount;
186 struct work_struct free_work;
187 };
188
189 #define pcidev_to_node(pcidev) pcibus_to_node(pcidev->bus)
190 #define ibdev_to_node(ibdev) pcidev_to_node(to_pci_dev(ibdev->dma_device))
191 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
192
193 /* bits for i_ack_flags */
194 #define IB_ACK_IN_FLIGHT 0
195 #define IB_ACK_REQUESTED 1
196
197 /* Magic WR_ID for ACKs */
198 #define RDS_IB_ACK_WR_ID (~(u64) 0)
199
200 struct rds_ib_statistics {
201 uint64_t s_ib_connect_raced;
202 uint64_t s_ib_listen_closed_stale;
203 uint64_t s_ib_tx_cq_call;
204 uint64_t s_ib_tx_cq_event;
205 uint64_t s_ib_tx_ring_full;
206 uint64_t s_ib_tx_throttle;
207 uint64_t s_ib_tx_sg_mapping_failure;
208 uint64_t s_ib_tx_stalled;
209 uint64_t s_ib_tx_credit_updates;
210 uint64_t s_ib_rx_cq_call;
211 uint64_t s_ib_rx_cq_event;
212 uint64_t s_ib_rx_ring_empty;
213 uint64_t s_ib_rx_refill_from_cq;
214 uint64_t s_ib_rx_refill_from_thread;
215 uint64_t s_ib_rx_alloc_limit;
216 uint64_t s_ib_rx_credit_updates;
217 uint64_t s_ib_ack_sent;
218 uint64_t s_ib_ack_send_failure;
219 uint64_t s_ib_ack_send_delayed;
220 uint64_t s_ib_ack_send_piggybacked;
221 uint64_t s_ib_ack_received;
222 uint64_t s_ib_rdma_mr_alloc;
223 uint64_t s_ib_rdma_mr_free;
224 uint64_t s_ib_rdma_mr_used;
225 uint64_t s_ib_rdma_mr_pool_flush;
226 uint64_t s_ib_rdma_mr_pool_wait;
227 uint64_t s_ib_rdma_mr_pool_depleted;
228 uint64_t s_ib_atomic_cswp;
229 uint64_t s_ib_atomic_fadd;
230 };
231
232 extern struct workqueue_struct *rds_ib_wq;
233
234 /*
235 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
236 * doesn't define it.
237 */
rds_ib_dma_sync_sg_for_cpu(struct ib_device * dev,struct scatterlist * sg,unsigned int sg_dma_len,int direction)238 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
239 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
240 {
241 unsigned int i;
242
243 for (i = 0; i < sg_dma_len; ++i) {
244 ib_dma_sync_single_for_cpu(dev,
245 ib_sg_dma_address(dev, &sg[i]),
246 ib_sg_dma_len(dev, &sg[i]),
247 direction);
248 }
249 }
250 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
251
rds_ib_dma_sync_sg_for_device(struct ib_device * dev,struct scatterlist * sg,unsigned int sg_dma_len,int direction)252 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
253 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
254 {
255 unsigned int i;
256
257 for (i = 0; i < sg_dma_len; ++i) {
258 ib_dma_sync_single_for_device(dev,
259 ib_sg_dma_address(dev, &sg[i]),
260 ib_sg_dma_len(dev, &sg[i]),
261 direction);
262 }
263 }
264 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
265
266
267 /* ib.c */
268 extern struct rds_transport rds_ib_transport;
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;
272
273 extern unsigned int fmr_message_size;
274 extern unsigned int rds_ib_retry_count;
275
276 extern spinlock_t ib_nodev_conns_lock;
277 extern struct list_head ib_nodev_conns;
278
279 /* ib_cm.c */
280 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
281 void rds_ib_conn_free(void *arg);
282 int rds_ib_conn_connect(struct rds_connection *conn);
283 void rds_ib_conn_shutdown(struct rds_connection *conn);
284 void rds_ib_state_change(struct sock *sk);
285 int rds_ib_listen_init(void);
286 void rds_ib_listen_stop(void);
287 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
288 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
289 struct rdma_cm_event *event);
290 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
291 void rds_ib_cm_connect_complete(struct rds_connection *conn,
292 struct rdma_cm_event *event);
293
294
295 #define rds_ib_conn_error(conn, fmt...) \
296 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
297
298 /* ib_rdma.c */
299 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
300 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
301 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
302 void rds_ib_destroy_nodev_conns(void);
303 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
304 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
305 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
306 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
307 struct rds_sock *rs, u32 *key_ret);
308 void rds_ib_sync_mr(void *trans_private, int dir);
309 void rds_ib_free_mr(void *trans_private, int invalidate);
310 void rds_ib_flush_mrs(void);
311
312 /* ib_recv.c */
313 int rds_ib_recv_init(void);
314 void rds_ib_recv_exit(void);
315 int rds_ib_recv(struct rds_connection *conn);
316 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic);
317 void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
318 void rds_ib_recv_refill(struct rds_connection *conn, int prefill);
319 void rds_ib_inc_free(struct rds_incoming *inc);
320 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
321 size_t size);
322 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
323 void rds_ib_recv_tasklet_fn(unsigned long data);
324 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
325 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
326 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
327 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
328 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
329 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
330
331 /* ib_ring.c */
332 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
333 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
334 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
335 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
336 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
337 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
338 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
339 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
340 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
341 extern wait_queue_head_t rds_ib_ring_empty_wait;
342
343 /* ib_send.c */
344 char *rds_ib_wc_status_str(enum ib_wc_status status);
345 void rds_ib_xmit_complete(struct rds_connection *conn);
346 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
347 unsigned int hdr_off, unsigned int sg, unsigned int off);
348 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
349 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
350 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
351 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
352 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
353 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
354 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
355 u32 *adv_credits, int need_posted, int max_posted);
356 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
357
358 /* ib_stats.c */
359 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
360 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
361 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
362 unsigned int avail);
363
364 /* ib_sysctl.c */
365 int rds_ib_sysctl_init(void);
366 void rds_ib_sysctl_exit(void);
367 extern unsigned long rds_ib_sysctl_max_send_wr;
368 extern unsigned long rds_ib_sysctl_max_recv_wr;
369 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
370 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
371 extern unsigned long rds_ib_sysctl_max_recv_allocation;
372 extern unsigned int rds_ib_sysctl_flow_control;
373
374 #endif
375