RDS/IB: Infiniband transport
[linux-2.6.git] / net / rds / ib.h
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 "rds.h"
7 #include "rdma_transport.h"
8
9 #define RDS_FMR_SIZE                    256
10 #define RDS_FMR_POOL_SIZE               4096
11
12 #define RDS_IB_MAX_SGE                  8
13 #define RDS_IB_RECV_SGE                 2
14
15 #define RDS_IB_DEFAULT_RECV_WR          1024
16 #define RDS_IB_DEFAULT_SEND_WR          256
17
18 #define RDS_IB_SUPPORTED_PROTOCOLS      0x00000003      /* minor versions supported */
19
20 extern struct list_head rds_ib_devices;
21
22 /*
23  * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
24  * try and minimize the amount of memory tied up both the device and
25  * socket receive queues.
26  */
27 /* page offset of the final full frag that fits in the page */
28 #define RDS_PAGE_LAST_OFF (((PAGE_SIZE  / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE)
29 struct rds_page_frag {
30         struct list_head        f_item;
31         struct page             *f_page;
32         unsigned long           f_offset;
33         dma_addr_t              f_mapped;
34 };
35
36 struct rds_ib_incoming {
37         struct list_head        ii_frags;
38         struct rds_incoming     ii_inc;
39 };
40
41 struct rds_ib_connect_private {
42         /* Add new fields at the end, and don't permute existing fields. */
43         __be32                  dp_saddr;
44         __be32                  dp_daddr;
45         u8                      dp_protocol_major;
46         u8                      dp_protocol_minor;
47         __be16                  dp_protocol_minor_mask; /* bitmask */
48         __be32                  dp_reserved1;
49         __be64                  dp_ack_seq;
50         __be32                  dp_credit;              /* non-zero enables flow ctl */
51 };
52
53 struct rds_ib_send_work {
54         struct rds_message      *s_rm;
55         struct rds_rdma_op      *s_op;
56         struct ib_send_wr       s_wr;
57         struct ib_sge           s_sge[RDS_IB_MAX_SGE];
58         unsigned long           s_queued;
59 };
60
61 struct rds_ib_recv_work {
62         struct rds_ib_incoming  *r_ibinc;
63         struct rds_page_frag    *r_frag;
64         struct ib_recv_wr       r_wr;
65         struct ib_sge           r_sge[2];
66 };
67
68 struct rds_ib_work_ring {
69         u32             w_nr;
70         u32             w_alloc_ptr;
71         u32             w_alloc_ctr;
72         u32             w_free_ptr;
73         atomic_t        w_free_ctr;
74 };
75
76 struct rds_ib_device;
77
78 struct rds_ib_connection {
79
80         struct list_head        ib_node;
81         struct rds_ib_device    *rds_ibdev;
82         struct rds_connection   *conn;
83
84         /* alphabet soup, IBTA style */
85         struct rdma_cm_id       *i_cm_id;
86         struct ib_pd            *i_pd;
87         struct ib_mr            *i_mr;
88         struct ib_cq            *i_send_cq;
89         struct ib_cq            *i_recv_cq;
90
91         /* tx */
92         struct rds_ib_work_ring i_send_ring;
93         struct rds_message      *i_rm;
94         struct rds_header       *i_send_hdrs;
95         u64                     i_send_hdrs_dma;
96         struct rds_ib_send_work *i_sends;
97
98         /* rx */
99         struct mutex            i_recv_mutex;
100         struct rds_ib_work_ring i_recv_ring;
101         struct rds_ib_incoming  *i_ibinc;
102         u32                     i_recv_data_rem;
103         struct rds_header       *i_recv_hdrs;
104         u64                     i_recv_hdrs_dma;
105         struct rds_ib_recv_work *i_recvs;
106         struct rds_page_frag    i_frag;
107         u64                     i_ack_recv;     /* last ACK received */
108
109         /* sending acks */
110         unsigned long           i_ack_flags;
111         u64                     i_ack_next;     /* next ACK to send */
112         struct rds_header       *i_ack;
113         struct ib_send_wr       i_ack_wr;
114         struct ib_sge           i_ack_sge;
115         u64                     i_ack_dma;
116         unsigned long           i_ack_queued;
117
118         /* Flow control related information
119          *
120          * Our algorithm uses a pair variables that we need to access
121          * atomically - one for the send credits, and one posted
122          * recv credits we need to transfer to remote.
123          * Rather than protect them using a slow spinlock, we put both into
124          * a single atomic_t and update it using cmpxchg
125          */
126         atomic_t                i_credits;
127
128         /* Protocol version specific information */
129         unsigned int            i_flowctl:1;    /* enable/disable flow ctl */
130
131         /* Batched completions */
132         unsigned int            i_unsignaled_wrs;
133         long                    i_unsignaled_bytes;
134 };
135
136 /* This assumes that atomic_t is at least 32 bits */
137 #define IB_GET_SEND_CREDITS(v)  ((v) & 0xffff)
138 #define IB_GET_POST_CREDITS(v)  ((v) >> 16)
139 #define IB_SET_SEND_CREDITS(v)  ((v) & 0xffff)
140 #define IB_SET_POST_CREDITS(v)  ((v) << 16)
141
142 struct rds_ib_ipaddr {
143         struct list_head        list;
144         __be32                  ipaddr;
145 };
146
147 struct rds_ib_device {
148         struct list_head        list;
149         struct list_head        ipaddr_list;
150         struct list_head        conn_list;
151         struct ib_device        *dev;
152         struct ib_pd            *pd;
153         struct ib_mr            *mr;
154         struct rds_ib_mr_pool   *mr_pool;
155         int                     fmr_page_shift;
156         int                     fmr_page_size;
157         u64                     fmr_page_mask;
158         unsigned int            fmr_max_remaps;
159         unsigned int            max_fmrs;
160         int                     max_sge;
161         unsigned int            max_wrs;
162         spinlock_t              spinlock;       /* protect the above */
163 };
164
165 /* bits for i_ack_flags */
166 #define IB_ACK_IN_FLIGHT        0
167 #define IB_ACK_REQUESTED        1
168
169 /* Magic WR_ID for ACKs */
170 #define RDS_IB_ACK_WR_ID        (~(u64) 0)
171
172 struct rds_ib_statistics {
173         uint64_t        s_ib_connect_raced;
174         uint64_t        s_ib_listen_closed_stale;
175         uint64_t        s_ib_tx_cq_call;
176         uint64_t        s_ib_tx_cq_event;
177         uint64_t        s_ib_tx_ring_full;
178         uint64_t        s_ib_tx_throttle;
179         uint64_t        s_ib_tx_sg_mapping_failure;
180         uint64_t        s_ib_tx_stalled;
181         uint64_t        s_ib_tx_credit_updates;
182         uint64_t        s_ib_rx_cq_call;
183         uint64_t        s_ib_rx_cq_event;
184         uint64_t        s_ib_rx_ring_empty;
185         uint64_t        s_ib_rx_refill_from_cq;
186         uint64_t        s_ib_rx_refill_from_thread;
187         uint64_t        s_ib_rx_alloc_limit;
188         uint64_t        s_ib_rx_credit_updates;
189         uint64_t        s_ib_ack_sent;
190         uint64_t        s_ib_ack_send_failure;
191         uint64_t        s_ib_ack_send_delayed;
192         uint64_t        s_ib_ack_send_piggybacked;
193         uint64_t        s_ib_ack_received;
194         uint64_t        s_ib_rdma_mr_alloc;
195         uint64_t        s_ib_rdma_mr_free;
196         uint64_t        s_ib_rdma_mr_used;
197         uint64_t        s_ib_rdma_mr_pool_flush;
198         uint64_t        s_ib_rdma_mr_pool_wait;
199         uint64_t        s_ib_rdma_mr_pool_depleted;
200 };
201
202 extern struct workqueue_struct *rds_ib_wq;
203
204 /*
205  * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
206  * doesn't define it.
207  */
208 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
209                 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
210 {
211         unsigned int i;
212
213         for (i = 0; i < sg_dma_len; ++i) {
214                 ib_dma_sync_single_for_cpu(dev,
215                                 ib_sg_dma_address(dev, &sg[i]),
216                                 ib_sg_dma_len(dev, &sg[i]),
217                                 direction);
218         }
219 }
220 #define ib_dma_sync_sg_for_cpu  rds_ib_dma_sync_sg_for_cpu
221
222 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
223                 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
224 {
225         unsigned int i;
226
227         for (i = 0; i < sg_dma_len; ++i) {
228                 ib_dma_sync_single_for_device(dev,
229                                 ib_sg_dma_address(dev, &sg[i]),
230                                 ib_sg_dma_len(dev, &sg[i]),
231                                 direction);
232         }
233 }
234 #define ib_dma_sync_sg_for_device       rds_ib_dma_sync_sg_for_device
235
236
237 /* ib.c */
238 extern struct rds_transport rds_ib_transport;
239 extern void rds_ib_add_one(struct ib_device *device);
240 extern void rds_ib_remove_one(struct ib_device *device);
241 extern struct ib_client rds_ib_client;
242
243 extern unsigned int fmr_pool_size;
244 extern unsigned int fmr_message_size;
245
246 extern spinlock_t ib_nodev_conns_lock;
247 extern struct list_head ib_nodev_conns;
248
249 /* ib_cm.c */
250 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
251 void rds_ib_conn_free(void *arg);
252 int rds_ib_conn_connect(struct rds_connection *conn);
253 void rds_ib_conn_shutdown(struct rds_connection *conn);
254 void rds_ib_state_change(struct sock *sk);
255 int __init rds_ib_listen_init(void);
256 void rds_ib_listen_stop(void);
257 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
258 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
259                              struct rdma_cm_event *event);
260 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
261 void rds_ib_cm_connect_complete(struct rds_connection *conn,
262                                 struct rdma_cm_event *event);
263
264
265 #define rds_ib_conn_error(conn, fmt...) \
266         __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
267
268 /* ib_rdma.c */
269 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
270 int rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
271 void rds_ib_remove_nodev_conns(void);
272 void rds_ib_remove_conns(struct rds_ib_device *rds_ibdev);
273 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
274 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
275 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
276 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
277                     struct rds_sock *rs, u32 *key_ret);
278 void rds_ib_sync_mr(void *trans_private, int dir);
279 void rds_ib_free_mr(void *trans_private, int invalidate);
280 void rds_ib_flush_mrs(void);
281
282 /* ib_recv.c */
283 int __init rds_ib_recv_init(void);
284 void rds_ib_recv_exit(void);
285 int rds_ib_recv(struct rds_connection *conn);
286 int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
287                        gfp_t page_gfp, int prefill);
288 void rds_ib_inc_purge(struct rds_incoming *inc);
289 void rds_ib_inc_free(struct rds_incoming *inc);
290 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
291                              size_t size);
292 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
293 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
294 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
295 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
296 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
297 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
298 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
299
300 /* ib_ring.c */
301 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
302 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
303 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
304 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
305 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
306 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
307 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
308 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
309 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
310 extern wait_queue_head_t rds_ib_ring_empty_wait;
311
312 /* ib_send.c */
313 void rds_ib_xmit_complete(struct rds_connection *conn);
314 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
315                 unsigned int hdr_off, unsigned int sg, unsigned int off);
316 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
317 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
318 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
319 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op);
320 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
321 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
322 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
323                              u32 *adv_credits, int need_posted);
324
325 /* ib_stats.c */
326 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
327 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
328 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
329                                     unsigned int avail);
330
331 /* ib_sysctl.c */
332 int __init rds_ib_sysctl_init(void);
333 void rds_ib_sysctl_exit(void);
334 extern unsigned long rds_ib_sysctl_max_send_wr;
335 extern unsigned long rds_ib_sysctl_max_recv_wr;
336 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
337 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
338 extern unsigned long rds_ib_sysctl_max_recv_allocation;
339 extern unsigned int rds_ib_sysctl_flow_control;
340 extern ctl_table rds_ib_sysctl_table[];
341
342 /*
343  * Helper functions for getting/setting the header and data SGEs in
344  * RDS packets (not RDMA)
345  */
346 static inline struct ib_sge *
347 rds_ib_header_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
348 {
349         return &sge[0];
350 }
351
352 static inline struct ib_sge *
353 rds_ib_data_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
354 {
355         return &sge[1];
356 }
357
358 static inline void rds_ib_set_64bit(u64 *ptr, u64 val)
359 {
360 #if BITS_PER_LONG == 64
361         *ptr = val;
362 #else
363         set_64bit(ptr, val);
364 #endif
365 }
366
367 #endif