1 #ifndef _RDS_IW_H
2 #define _RDS_IW_H
3
4 #include <linux/interrupt.h>
5 #include <rdma/ib_verbs.h>
6 #include <rdma/rdma_cm.h>
7 #include "rds.h"
8 #include "rdma_transport.h"
9
10 #define RDS_FASTREG_SIZE 20
11 #define RDS_FASTREG_POOL_SIZE 2048
12
13 #define RDS_IW_MAX_SGE 8
14 #define RDS_IW_RECV_SGE 2
15
16 #define RDS_IW_DEFAULT_RECV_WR 1024
17 #define RDS_IW_DEFAULT_SEND_WR 256
18
19 #define RDS_IW_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
20
21 extern struct list_head rds_iw_devices;
22
23 /*
24 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
25 * try and minimize the amount of memory tied up both the device and
26 * socket receive queues.
27 */
28 /* page offset of the final full frag that fits in the page */
29 #define RDS_PAGE_LAST_OFF (((PAGE_SIZE / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE)
30 struct rds_page_frag {
31 struct list_head f_item;
32 struct page *f_page;
33 unsigned long f_offset;
34 dma_addr_t f_mapped;
35 };
36
37 struct rds_iw_incoming {
38 struct list_head ii_frags;
39 struct rds_incoming ii_inc;
40 };
41
42 struct rds_iw_connect_private {
43 /* Add new fields at the end, and don't permute existing fields. */
44 __be32 dp_saddr;
45 __be32 dp_daddr;
46 u8 dp_protocol_major;
47 u8 dp_protocol_minor;
48 __be16 dp_protocol_minor_mask; /* bitmask */
49 __be32 dp_reserved1;
50 __be64 dp_ack_seq;
51 __be32 dp_credit; /* non-zero enables flow ctl */
52 };
53
54 struct rds_iw_scatterlist {
55 struct scatterlist *list;
56 unsigned int len;
57 int dma_len;
58 unsigned int dma_npages;
59 unsigned int bytes;
60 };
61
62 struct rds_iw_mapping {
63 spinlock_t m_lock; /* protect the mapping struct */
64 struct list_head m_list;
65 struct rds_iw_mr *m_mr;
66 uint32_t m_rkey;
67 struct rds_iw_scatterlist m_sg;
68 };
69
70 struct rds_iw_send_work {
71 struct rds_message *s_rm;
72
73 /* We should really put these into a union: */
74 struct rm_rdma_op *s_op;
75 struct rds_iw_mapping *s_mapping;
76 struct ib_mr *s_mr;
77 unsigned char s_remap_count;
78
79 union {
80 struct ib_send_wr s_send_wr;
81 struct ib_rdma_wr s_rdma_wr;
82 struct ib_reg_wr s_reg_wr;
83 };
84 struct ib_sge s_sge[RDS_IW_MAX_SGE];
85 unsigned long s_queued;
86 };
87
88 struct rds_iw_recv_work {
89 struct rds_iw_incoming *r_iwinc;
90 struct rds_page_frag *r_frag;
91 struct ib_recv_wr r_wr;
92 struct ib_sge r_sge[2];
93 };
94
95 struct rds_iw_work_ring {
96 u32 w_nr;
97 u32 w_alloc_ptr;
98 u32 w_alloc_ctr;
99 u32 w_free_ptr;
100 atomic_t w_free_ctr;
101 };
102
103 struct rds_iw_device;
104
105 struct rds_iw_connection {
106
107 struct list_head iw_node;
108 struct rds_iw_device *rds_iwdev;
109 struct rds_connection *conn;
110
111 /* alphabet soup, IBTA style */
112 struct rdma_cm_id *i_cm_id;
113 struct ib_pd *i_pd;
114 struct ib_mr *i_mr;
115 struct ib_cq *i_send_cq;
116 struct ib_cq *i_recv_cq;
117
118 /* tx */
119 struct rds_iw_work_ring i_send_ring;
120 struct rds_message *i_rm;
121 struct rds_header *i_send_hdrs;
122 u64 i_send_hdrs_dma;
123 struct rds_iw_send_work *i_sends;
124
125 /* rx */
126 struct tasklet_struct i_recv_tasklet;
127 struct mutex i_recv_mutex;
128 struct rds_iw_work_ring i_recv_ring;
129 struct rds_iw_incoming *i_iwinc;
130 u32 i_recv_data_rem;
131 struct rds_header *i_recv_hdrs;
132 u64 i_recv_hdrs_dma;
133 struct rds_iw_recv_work *i_recvs;
134 struct rds_page_frag i_frag;
135 u64 i_ack_recv; /* last ACK received */
136
137 /* sending acks */
138 unsigned long i_ack_flags;
139 #ifdef KERNEL_HAS_ATOMIC64
140 atomic64_t i_ack_next; /* next ACK to send */
141 #else
142 spinlock_t i_ack_lock; /* protect i_ack_next */
143 u64 i_ack_next; /* next ACK to send */
144 #endif
145 struct rds_header *i_ack;
146 struct ib_send_wr i_ack_wr;
147 struct ib_sge i_ack_sge;
148 u64 i_ack_dma;
149 unsigned long i_ack_queued;
150
151 /* Flow control related information
152 *
153 * Our algorithm uses a pair variables that we need to access
154 * atomically - one for the send credits, and one posted
155 * recv credits we need to transfer to remote.
156 * Rather than protect them using a slow spinlock, we put both into
157 * a single atomic_t and update it using cmpxchg
158 */
159 atomic_t i_credits;
160
161 /* Protocol version specific information */
162 unsigned int i_flowctl:1; /* enable/disable flow ctl */
163 unsigned int i_dma_local_lkey:1;
164 unsigned int i_fastreg_posted:1; /* fastreg posted on this connection */
165 /* Batched completions */
166 unsigned int i_unsignaled_wrs;
167 long i_unsignaled_bytes;
168 };
169
170 /* This assumes that atomic_t is at least 32 bits */
171 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
172 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
173 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
174 #define IB_SET_POST_CREDITS(v) ((v) << 16)
175
176 struct rds_iw_cm_id {
177 struct list_head list;
178 struct rdma_cm_id *cm_id;
179 };
180
181 struct rds_iw_device {
182 struct list_head list;
183 struct list_head cm_id_list;
184 struct list_head conn_list;
185 struct ib_device *dev;
186 struct ib_pd *pd;
187 struct ib_mr *mr;
188 struct rds_iw_mr_pool *mr_pool;
189 int max_sge;
190 unsigned int max_wrs;
191 unsigned int dma_local_lkey:1;
192 spinlock_t spinlock; /* protect the above */
193 };
194
195 /* bits for i_ack_flags */
196 #define IB_ACK_IN_FLIGHT 0
197 #define IB_ACK_REQUESTED 1
198
199 /* Magic WR_ID for ACKs */
200 #define RDS_IW_ACK_WR_ID ((u64)0xffffffffffffffffULL)
201 #define RDS_IW_REG_WR_ID ((u64)0xefefefefefefefefULL)
202 #define RDS_IW_LOCAL_INV_WR_ID ((u64)0xdfdfdfdfdfdfdfdfULL)
203
204 struct rds_iw_statistics {
205 uint64_t s_iw_connect_raced;
206 uint64_t s_iw_listen_closed_stale;
207 uint64_t s_iw_tx_cq_call;
208 uint64_t s_iw_tx_cq_event;
209 uint64_t s_iw_tx_ring_full;
210 uint64_t s_iw_tx_throttle;
211 uint64_t s_iw_tx_sg_mapping_failure;
212 uint64_t s_iw_tx_stalled;
213 uint64_t s_iw_tx_credit_updates;
214 uint64_t s_iw_rx_cq_call;
215 uint64_t s_iw_rx_cq_event;
216 uint64_t s_iw_rx_ring_empty;
217 uint64_t s_iw_rx_refill_from_cq;
218 uint64_t s_iw_rx_refill_from_thread;
219 uint64_t s_iw_rx_alloc_limit;
220 uint64_t s_iw_rx_credit_updates;
221 uint64_t s_iw_ack_sent;
222 uint64_t s_iw_ack_send_failure;
223 uint64_t s_iw_ack_send_delayed;
224 uint64_t s_iw_ack_send_piggybacked;
225 uint64_t s_iw_ack_received;
226 uint64_t s_iw_rdma_mr_alloc;
227 uint64_t s_iw_rdma_mr_free;
228 uint64_t s_iw_rdma_mr_used;
229 uint64_t s_iw_rdma_mr_pool_flush;
230 uint64_t s_iw_rdma_mr_pool_wait;
231 uint64_t s_iw_rdma_mr_pool_depleted;
232 };
233
234 extern struct workqueue_struct *rds_iw_wq;
235
236 /*
237 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
238 * doesn't define it.
239 */
rds_iw_dma_sync_sg_for_cpu(struct ib_device * dev,struct scatterlist * sg,unsigned int sg_dma_len,int direction)240 static inline void rds_iw_dma_sync_sg_for_cpu(struct ib_device *dev,
241 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
242 {
243 unsigned int i;
244
245 for (i = 0; i < sg_dma_len; ++i) {
246 ib_dma_sync_single_for_cpu(dev,
247 ib_sg_dma_address(dev, &sg[i]),
248 ib_sg_dma_len(dev, &sg[i]),
249 direction);
250 }
251 }
252 #define ib_dma_sync_sg_for_cpu rds_iw_dma_sync_sg_for_cpu
253
rds_iw_dma_sync_sg_for_device(struct ib_device * dev,struct scatterlist * sg,unsigned int sg_dma_len,int direction)254 static inline void rds_iw_dma_sync_sg_for_device(struct ib_device *dev,
255 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
256 {
257 unsigned int i;
258
259 for (i = 0; i < sg_dma_len; ++i) {
260 ib_dma_sync_single_for_device(dev,
261 ib_sg_dma_address(dev, &sg[i]),
262 ib_sg_dma_len(dev, &sg[i]),
263 direction);
264 }
265 }
266 #define ib_dma_sync_sg_for_device rds_iw_dma_sync_sg_for_device
267
rds_iw_local_dma_lkey(struct rds_iw_connection * ic)268 static inline u32 rds_iw_local_dma_lkey(struct rds_iw_connection *ic)
269 {
270 return ic->i_dma_local_lkey ? ic->i_cm_id->device->local_dma_lkey : ic->i_mr->lkey;
271 }
272
273 /* ib.c */
274 extern struct rds_transport rds_iw_transport;
275 extern struct ib_client rds_iw_client;
276
277 extern unsigned int fastreg_pool_size;
278 extern unsigned int fastreg_message_size;
279
280 extern spinlock_t iw_nodev_conns_lock;
281 extern struct list_head iw_nodev_conns;
282
283 /* ib_cm.c */
284 int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp);
285 void rds_iw_conn_free(void *arg);
286 int rds_iw_conn_connect(struct rds_connection *conn);
287 void rds_iw_conn_shutdown(struct rds_connection *conn);
288 void rds_iw_state_change(struct sock *sk);
289 int rds_iw_listen_init(void);
290 void rds_iw_listen_stop(void);
291 void __rds_iw_conn_error(struct rds_connection *conn, const char *, ...);
292 int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
293 struct rdma_cm_event *event);
294 int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id);
295 void rds_iw_cm_connect_complete(struct rds_connection *conn,
296 struct rdma_cm_event *event);
297
298
299 #define rds_iw_conn_error(conn, fmt...) \
300 __rds_iw_conn_error(conn, KERN_WARNING "RDS/IW: " fmt)
301
302 /* ib_rdma.c */
303 int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
304 void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
305 void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
306 void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock);
rds_iw_destroy_nodev_conns(void)307 static inline void rds_iw_destroy_nodev_conns(void)
308 {
309 __rds_iw_destroy_conns(&iw_nodev_conns, &iw_nodev_conns_lock);
310 }
rds_iw_destroy_conns(struct rds_iw_device * rds_iwdev)311 static inline void rds_iw_destroy_conns(struct rds_iw_device *rds_iwdev)
312 {
313 __rds_iw_destroy_conns(&rds_iwdev->conn_list, &rds_iwdev->spinlock);
314 }
315 struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *);
316 void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo);
317 void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *);
318 void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
319 struct rds_sock *rs, u32 *key_ret);
320 void rds_iw_sync_mr(void *trans_private, int dir);
321 void rds_iw_free_mr(void *trans_private, int invalidate);
322 void rds_iw_flush_mrs(void);
323
324 /* ib_recv.c */
325 int rds_iw_recv_init(void);
326 void rds_iw_recv_exit(void);
327 int rds_iw_recv(struct rds_connection *conn);
328 int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
329 gfp_t page_gfp, int prefill);
330 void rds_iw_inc_free(struct rds_incoming *inc);
331 int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
332 void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context);
333 void rds_iw_recv_tasklet_fn(unsigned long data);
334 void rds_iw_recv_init_ring(struct rds_iw_connection *ic);
335 void rds_iw_recv_clear_ring(struct rds_iw_connection *ic);
336 void rds_iw_recv_init_ack(struct rds_iw_connection *ic);
337 void rds_iw_attempt_ack(struct rds_iw_connection *ic);
338 void rds_iw_ack_send_complete(struct rds_iw_connection *ic);
339 u64 rds_iw_piggyb_ack(struct rds_iw_connection *ic);
340
341 /* ib_ring.c */
342 void rds_iw_ring_init(struct rds_iw_work_ring *ring, u32 nr);
343 void rds_iw_ring_resize(struct rds_iw_work_ring *ring, u32 nr);
344 u32 rds_iw_ring_alloc(struct rds_iw_work_ring *ring, u32 val, u32 *pos);
345 void rds_iw_ring_free(struct rds_iw_work_ring *ring, u32 val);
346 void rds_iw_ring_unalloc(struct rds_iw_work_ring *ring, u32 val);
347 int rds_iw_ring_empty(struct rds_iw_work_ring *ring);
348 int rds_iw_ring_low(struct rds_iw_work_ring *ring);
349 u32 rds_iw_ring_oldest(struct rds_iw_work_ring *ring);
350 u32 rds_iw_ring_completed(struct rds_iw_work_ring *ring, u32 wr_id, u32 oldest);
351 extern wait_queue_head_t rds_iw_ring_empty_wait;
352
353 /* ib_send.c */
354 void rds_iw_xmit_complete(struct rds_connection *conn);
355 int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
356 unsigned int hdr_off, unsigned int sg, unsigned int off);
357 void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context);
358 void rds_iw_send_init_ring(struct rds_iw_connection *ic);
359 void rds_iw_send_clear_ring(struct rds_iw_connection *ic);
360 int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
361 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits);
362 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted);
363 int rds_iw_send_grab_credits(struct rds_iw_connection *ic, u32 wanted,
364 u32 *adv_credits, int need_posted, int max_posted);
365
366 /* ib_stats.c */
367 DECLARE_PER_CPU(struct rds_iw_statistics, rds_iw_stats);
368 #define rds_iw_stats_inc(member) rds_stats_inc_which(rds_iw_stats, member)
369 unsigned int rds_iw_stats_info_copy(struct rds_info_iterator *iter,
370 unsigned int avail);
371
372 /* ib_sysctl.c */
373 int rds_iw_sysctl_init(void);
374 void rds_iw_sysctl_exit(void);
375 extern unsigned long rds_iw_sysctl_max_send_wr;
376 extern unsigned long rds_iw_sysctl_max_recv_wr;
377 extern unsigned long rds_iw_sysctl_max_unsig_wrs;
378 extern unsigned long rds_iw_sysctl_max_unsig_bytes;
379 extern unsigned long rds_iw_sysctl_max_recv_allocation;
380 extern unsigned int rds_iw_sysctl_flow_control;
381
382 /*
383 * Helper functions for getting/setting the header and data SGEs in
384 * RDS packets (not RDMA)
385 */
386 static inline struct ib_sge *
rds_iw_header_sge(struct rds_iw_connection * ic,struct ib_sge * sge)387 rds_iw_header_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
388 {
389 return &sge[0];
390 }
391
392 static inline struct ib_sge *
rds_iw_data_sge(struct rds_iw_connection * ic,struct ib_sge * sge)393 rds_iw_data_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
394 {
395 return &sge[1];
396 }
397
398 #endif
399