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 struct ib_fast_reg_page_list *s_page_list;
78 unsigned char s_remap_count;
79
80 struct ib_send_wr s_wr;
81 struct ib_sge s_sge[RDS_IW_MAX_SGE];
82 unsigned long s_queued;
83 };
84
85 struct rds_iw_recv_work {
86 struct rds_iw_incoming *r_iwinc;
87 struct rds_page_frag *r_frag;
88 struct ib_recv_wr r_wr;
89 struct ib_sge r_sge[2];
90 };
91
92 struct rds_iw_work_ring {
93 u32 w_nr;
94 u32 w_alloc_ptr;
95 u32 w_alloc_ctr;
96 u32 w_free_ptr;
97 atomic_t w_free_ctr;
98 };
99
100 struct rds_iw_device;
101
102 struct rds_iw_connection {
103
104 struct list_head iw_node;
105 struct rds_iw_device *rds_iwdev;
106 struct rds_connection *conn;
107
108 /* alphabet soup, IBTA style */
109 struct rdma_cm_id *i_cm_id;
110 struct ib_pd *i_pd;
111 struct ib_mr *i_mr;
112 struct ib_cq *i_send_cq;
113 struct ib_cq *i_recv_cq;
114
115 /* tx */
116 struct rds_iw_work_ring i_send_ring;
117 struct rds_message *i_rm;
118 struct rds_header *i_send_hdrs;
119 u64 i_send_hdrs_dma;
120 struct rds_iw_send_work *i_sends;
121
122 /* rx */
123 struct tasklet_struct i_recv_tasklet;
124 struct mutex i_recv_mutex;
125 struct rds_iw_work_ring i_recv_ring;
126 struct rds_iw_incoming *i_iwinc;
127 u32 i_recv_data_rem;
128 struct rds_header *i_recv_hdrs;
129 u64 i_recv_hdrs_dma;
130 struct rds_iw_recv_work *i_recvs;
131 struct rds_page_frag i_frag;
132 u64 i_ack_recv; /* last ACK received */
133
134 /* sending acks */
135 unsigned long i_ack_flags;
136 #ifdef KERNEL_HAS_ATOMIC64
137 atomic64_t i_ack_next; /* next ACK to send */
138 #else
139 spinlock_t i_ack_lock; /* protect i_ack_next */
140 u64 i_ack_next; /* next ACK to send */
141 #endif
142 struct rds_header *i_ack;
143 struct ib_send_wr i_ack_wr;
144 struct ib_sge i_ack_sge;
145 u64 i_ack_dma;
146 unsigned long i_ack_queued;
147
148 /* Flow control related information
149 *
150 * Our algorithm uses a pair variables that we need to access
151 * atomically - one for the send credits, and one posted
152 * recv credits we need to transfer to remote.
153 * Rather than protect them using a slow spinlock, we put both into
154 * a single atomic_t and update it using cmpxchg
155 */
156 atomic_t i_credits;
157
158 /* Protocol version specific information */
159 unsigned int i_flowctl:1; /* enable/disable flow ctl */
160 unsigned int i_dma_local_lkey:1;
161 unsigned int i_fastreg_posted:1; /* fastreg posted on this connection */
162 /* Batched completions */
163 unsigned int i_unsignaled_wrs;
164 long i_unsignaled_bytes;
165 };
166
167 /* This assumes that atomic_t is at least 32 bits */
168 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
169 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
170 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
171 #define IB_SET_POST_CREDITS(v) ((v) << 16)
172
173 struct rds_iw_cm_id {
174 struct list_head list;
175 struct rdma_cm_id *cm_id;
176 };
177
178 struct rds_iw_device {
179 struct list_head list;
180 struct list_head cm_id_list;
181 struct list_head conn_list;
182 struct ib_device *dev;
183 struct ib_pd *pd;
184 struct ib_mr *mr;
185 struct rds_iw_mr_pool *mr_pool;
186 int max_sge;
187 unsigned int max_wrs;
188 unsigned int dma_local_lkey:1;
189 spinlock_t spinlock; /* protect the above */
190 };
191
192 /* bits for i_ack_flags */
193 #define IB_ACK_IN_FLIGHT 0
194 #define IB_ACK_REQUESTED 1
195
196 /* Magic WR_ID for ACKs */
197 #define RDS_IW_ACK_WR_ID ((u64)0xffffffffffffffffULL)
198 #define RDS_IW_FAST_REG_WR_ID ((u64)0xefefefefefefefefULL)
199 #define RDS_IW_LOCAL_INV_WR_ID ((u64)0xdfdfdfdfdfdfdfdfULL)
200
201 struct rds_iw_statistics {
202 uint64_t s_iw_connect_raced;
203 uint64_t s_iw_listen_closed_stale;
204 uint64_t s_iw_tx_cq_call;
205 uint64_t s_iw_tx_cq_event;
206 uint64_t s_iw_tx_ring_full;
207 uint64_t s_iw_tx_throttle;
208 uint64_t s_iw_tx_sg_mapping_failure;
209 uint64_t s_iw_tx_stalled;
210 uint64_t s_iw_tx_credit_updates;
211 uint64_t s_iw_rx_cq_call;
212 uint64_t s_iw_rx_cq_event;
213 uint64_t s_iw_rx_ring_empty;
214 uint64_t s_iw_rx_refill_from_cq;
215 uint64_t s_iw_rx_refill_from_thread;
216 uint64_t s_iw_rx_alloc_limit;
217 uint64_t s_iw_rx_credit_updates;
218 uint64_t s_iw_ack_sent;
219 uint64_t s_iw_ack_send_failure;
220 uint64_t s_iw_ack_send_delayed;
221 uint64_t s_iw_ack_send_piggybacked;
222 uint64_t s_iw_ack_received;
223 uint64_t s_iw_rdma_mr_alloc;
224 uint64_t s_iw_rdma_mr_free;
225 uint64_t s_iw_rdma_mr_used;
226 uint64_t s_iw_rdma_mr_pool_flush;
227 uint64_t s_iw_rdma_mr_pool_wait;
228 uint64_t s_iw_rdma_mr_pool_depleted;
229 };
230
231 extern struct workqueue_struct *rds_iw_wq;
232
233 /*
234 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
235 * doesn't define it.
236 */
rds_iw_dma_sync_sg_for_cpu(struct ib_device * dev,struct scatterlist * sg,unsigned int sg_dma_len,int direction)237 static inline void rds_iw_dma_sync_sg_for_cpu(struct ib_device *dev,
238 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
239 {
240 unsigned int i;
241
242 for (i = 0; i < sg_dma_len; ++i) {
243 ib_dma_sync_single_for_cpu(dev,
244 ib_sg_dma_address(dev, &sg[i]),
245 ib_sg_dma_len(dev, &sg[i]),
246 direction);
247 }
248 }
249 #define ib_dma_sync_sg_for_cpu rds_iw_dma_sync_sg_for_cpu
250
rds_iw_dma_sync_sg_for_device(struct ib_device * dev,struct scatterlist * sg,unsigned int sg_dma_len,int direction)251 static inline void rds_iw_dma_sync_sg_for_device(struct ib_device *dev,
252 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
253 {
254 unsigned int i;
255
256 for (i = 0; i < sg_dma_len; ++i) {
257 ib_dma_sync_single_for_device(dev,
258 ib_sg_dma_address(dev, &sg[i]),
259 ib_sg_dma_len(dev, &sg[i]),
260 direction);
261 }
262 }
263 #define ib_dma_sync_sg_for_device rds_iw_dma_sync_sg_for_device
264
rds_iw_local_dma_lkey(struct rds_iw_connection * ic)265 static inline u32 rds_iw_local_dma_lkey(struct rds_iw_connection *ic)
266 {
267 return ic->i_dma_local_lkey ? ic->i_cm_id->device->local_dma_lkey : ic->i_mr->lkey;
268 }
269
270 /* ib.c */
271 extern struct rds_transport rds_iw_transport;
272 extern struct ib_client rds_iw_client;
273
274 extern unsigned int fastreg_pool_size;
275 extern unsigned int fastreg_message_size;
276
277 extern spinlock_t iw_nodev_conns_lock;
278 extern struct list_head iw_nodev_conns;
279
280 /* ib_cm.c */
281 int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp);
282 void rds_iw_conn_free(void *arg);
283 int rds_iw_conn_connect(struct rds_connection *conn);
284 void rds_iw_conn_shutdown(struct rds_connection *conn);
285 void rds_iw_state_change(struct sock *sk);
286 int rds_iw_listen_init(void);
287 void rds_iw_listen_stop(void);
288 void __rds_iw_conn_error(struct rds_connection *conn, const char *, ...);
289 int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
290 struct rdma_cm_event *event);
291 int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id);
292 void rds_iw_cm_connect_complete(struct rds_connection *conn,
293 struct rdma_cm_event *event);
294
295
296 #define rds_iw_conn_error(conn, fmt...) \
297 __rds_iw_conn_error(conn, KERN_WARNING "RDS/IW: " fmt)
298
299 /* ib_rdma.c */
300 int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
301 void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
302 void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
303 void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock);
rds_iw_destroy_nodev_conns(void)304 static inline void rds_iw_destroy_nodev_conns(void)
305 {
306 __rds_iw_destroy_conns(&iw_nodev_conns, &iw_nodev_conns_lock);
307 }
rds_iw_destroy_conns(struct rds_iw_device * rds_iwdev)308 static inline void rds_iw_destroy_conns(struct rds_iw_device *rds_iwdev)
309 {
310 __rds_iw_destroy_conns(&rds_iwdev->conn_list, &rds_iwdev->spinlock);
311 }
312 struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *);
313 void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo);
314 void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *);
315 void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
316 struct rds_sock *rs, u32 *key_ret);
317 void rds_iw_sync_mr(void *trans_private, int dir);
318 void rds_iw_free_mr(void *trans_private, int invalidate);
319 void rds_iw_flush_mrs(void);
320
321 /* ib_recv.c */
322 int rds_iw_recv_init(void);
323 void rds_iw_recv_exit(void);
324 int rds_iw_recv(struct rds_connection *conn);
325 int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
326 gfp_t page_gfp, int prefill);
327 void rds_iw_inc_free(struct rds_incoming *inc);
328 int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
329 size_t size);
330 void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context);
331 void rds_iw_recv_tasklet_fn(unsigned long data);
332 void rds_iw_recv_init_ring(struct rds_iw_connection *ic);
333 void rds_iw_recv_clear_ring(struct rds_iw_connection *ic);
334 void rds_iw_recv_init_ack(struct rds_iw_connection *ic);
335 void rds_iw_attempt_ack(struct rds_iw_connection *ic);
336 void rds_iw_ack_send_complete(struct rds_iw_connection *ic);
337 u64 rds_iw_piggyb_ack(struct rds_iw_connection *ic);
338
339 /* ib_ring.c */
340 void rds_iw_ring_init(struct rds_iw_work_ring *ring, u32 nr);
341 void rds_iw_ring_resize(struct rds_iw_work_ring *ring, u32 nr);
342 u32 rds_iw_ring_alloc(struct rds_iw_work_ring *ring, u32 val, u32 *pos);
343 void rds_iw_ring_free(struct rds_iw_work_ring *ring, u32 val);
344 void rds_iw_ring_unalloc(struct rds_iw_work_ring *ring, u32 val);
345 int rds_iw_ring_empty(struct rds_iw_work_ring *ring);
346 int rds_iw_ring_low(struct rds_iw_work_ring *ring);
347 u32 rds_iw_ring_oldest(struct rds_iw_work_ring *ring);
348 u32 rds_iw_ring_completed(struct rds_iw_work_ring *ring, u32 wr_id, u32 oldest);
349 extern wait_queue_head_t rds_iw_ring_empty_wait;
350
351 /* ib_send.c */
352 void rds_iw_xmit_complete(struct rds_connection *conn);
353 int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
354 unsigned int hdr_off, unsigned int sg, unsigned int off);
355 void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context);
356 void rds_iw_send_init_ring(struct rds_iw_connection *ic);
357 void rds_iw_send_clear_ring(struct rds_iw_connection *ic);
358 int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
359 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits);
360 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted);
361 int rds_iw_send_grab_credits(struct rds_iw_connection *ic, u32 wanted,
362 u32 *adv_credits, int need_posted, int max_posted);
363
364 /* ib_stats.c */
365 DECLARE_PER_CPU(struct rds_iw_statistics, rds_iw_stats);
366 #define rds_iw_stats_inc(member) rds_stats_inc_which(rds_iw_stats, member)
367 unsigned int rds_iw_stats_info_copy(struct rds_info_iterator *iter,
368 unsigned int avail);
369
370 /* ib_sysctl.c */
371 int rds_iw_sysctl_init(void);
372 void rds_iw_sysctl_exit(void);
373 extern unsigned long rds_iw_sysctl_max_send_wr;
374 extern unsigned long rds_iw_sysctl_max_recv_wr;
375 extern unsigned long rds_iw_sysctl_max_unsig_wrs;
376 extern unsigned long rds_iw_sysctl_max_unsig_bytes;
377 extern unsigned long rds_iw_sysctl_max_recv_allocation;
378 extern unsigned int rds_iw_sysctl_flow_control;
379
380 /*
381 * Helper functions for getting/setting the header and data SGEs in
382 * RDS packets (not RDMA)
383 */
384 static inline struct ib_sge *
rds_iw_header_sge(struct rds_iw_connection * ic,struct ib_sge * sge)385 rds_iw_header_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
386 {
387 return &sge[0];
388 }
389
390 static inline struct ib_sge *
rds_iw_data_sge(struct rds_iw_connection * ic,struct ib_sge * sge)391 rds_iw_data_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
392 {
393 return &sge[1];
394 }
395
396 #endif
397