blob: 0715dde323e77c88a56f50e591bb065a900efbca [file] [log] [blame]
Andy Groverfcd8b7c2009-02-24 15:30:36 +00001#ifndef _RDS_IW_H
2#define _RDS_IW_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_FASTREG_SIZE 20
10#define RDS_FASTREG_POOL_SIZE 2048
11
12#define RDS_IW_MAX_SGE 8
13#define RDS_IW_RECV_SGE 2
14
15#define RDS_IW_DEFAULT_RECV_WR 1024
16#define RDS_IW_DEFAULT_SEND_WR 256
17
18#define RDS_IW_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
19
20extern struct list_head rds_iw_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)
29struct 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
36struct rds_iw_incoming {
37 struct list_head ii_frags;
38 struct rds_incoming ii_inc;
39};
40
41struct rds_iw_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
53struct rds_iw_scatterlist {
54 struct scatterlist *list;
55 unsigned int len;
56 int dma_len;
57 unsigned int dma_npages;
58 unsigned int bytes;
59};
60
61struct rds_iw_mapping {
62 spinlock_t m_lock; /* protect the mapping struct */
63 struct list_head m_list;
64 struct rds_iw_mr *m_mr;
65 uint32_t m_rkey;
66 struct rds_iw_scatterlist m_sg;
67};
68
69struct rds_iw_send_work {
70 struct rds_message *s_rm;
71
72 /* We should really put these into a union: */
73 struct rds_rdma_op *s_op;
74 struct rds_iw_mapping *s_mapping;
75 struct ib_mr *s_mr;
76 struct ib_fast_reg_page_list *s_page_list;
77 unsigned char s_remap_count;
78
79 struct ib_send_wr s_wr;
80 struct ib_sge s_sge[RDS_IW_MAX_SGE];
81 unsigned long s_queued;
82};
83
84struct rds_iw_recv_work {
85 struct rds_iw_incoming *r_iwinc;
86 struct rds_page_frag *r_frag;
87 struct ib_recv_wr r_wr;
88 struct ib_sge r_sge[2];
89};
90
91struct rds_iw_work_ring {
92 u32 w_nr;
93 u32 w_alloc_ptr;
94 u32 w_alloc_ctr;
95 u32 w_free_ptr;
96 atomic_t w_free_ctr;
97};
98
99struct rds_iw_device;
100
101struct rds_iw_connection {
102
103 struct list_head iw_node;
104 struct rds_iw_device *rds_iwdev;
105 struct rds_connection *conn;
106
107 /* alphabet soup, IBTA style */
108 struct rdma_cm_id *i_cm_id;
109 struct ib_pd *i_pd;
110 struct ib_mr *i_mr;
111 struct ib_cq *i_send_cq;
112 struct ib_cq *i_recv_cq;
113
114 /* tx */
115 struct rds_iw_work_ring i_send_ring;
116 struct rds_message *i_rm;
117 struct rds_header *i_send_hdrs;
118 u64 i_send_hdrs_dma;
119 struct rds_iw_send_work *i_sends;
120
121 /* rx */
122 struct mutex i_recv_mutex;
123 struct rds_iw_work_ring i_recv_ring;
124 struct rds_iw_incoming *i_iwinc;
125 u32 i_recv_data_rem;
126 struct rds_header *i_recv_hdrs;
127 u64 i_recv_hdrs_dma;
128 struct rds_iw_recv_work *i_recvs;
129 struct rds_page_frag i_frag;
130 u64 i_ack_recv; /* last ACK received */
131
132 /* sending acks */
133 unsigned long i_ack_flags;
Andy Grover8cbd9602009-04-01 08:20:20 +0000134#ifdef KERNEL_HAS_ATOMIC64
135 atomic64_t i_ack_next; /* next ACK to send */
136#else
137 spinlock_t i_ack_lock; /* protect i_ack_next */
Andy Groverfcd8b7c2009-02-24 15:30:36 +0000138 u64 i_ack_next; /* next ACK to send */
Andy Grover8cbd9602009-04-01 08:20:20 +0000139#endif
Andy Groverfcd8b7c2009-02-24 15:30:36 +0000140 struct rds_header *i_ack;
141 struct ib_send_wr i_ack_wr;
142 struct ib_sge i_ack_sge;
143 u64 i_ack_dma;
144 unsigned long i_ack_queued;
145
146 /* Flow control related information
147 *
148 * Our algorithm uses a pair variables that we need to access
149 * atomically - one for the send credits, and one posted
150 * recv credits we need to transfer to remote.
151 * Rather than protect them using a slow spinlock, we put both into
152 * a single atomic_t and update it using cmpxchg
153 */
154 atomic_t i_credits;
155
156 /* Protocol version specific information */
157 unsigned int i_flowctl:1; /* enable/disable flow ctl */
158 unsigned int i_dma_local_lkey:1;
159 unsigned int i_fastreg_posted:1; /* fastreg posted on this connection */
160 /* Batched completions */
161 unsigned int i_unsignaled_wrs;
162 long i_unsignaled_bytes;
163};
164
165/* This assumes that atomic_t is at least 32 bits */
166#define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
167#define IB_GET_POST_CREDITS(v) ((v) >> 16)
168#define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
169#define IB_SET_POST_CREDITS(v) ((v) << 16)
170
171struct rds_iw_cm_id {
172 struct list_head list;
173 struct rdma_cm_id *cm_id;
174};
175
176struct rds_iw_device {
177 struct list_head list;
178 struct list_head cm_id_list;
179 struct list_head conn_list;
180 struct ib_device *dev;
181 struct ib_pd *pd;
182 struct ib_mr *mr;
183 struct rds_iw_mr_pool *mr_pool;
184 int page_shift;
185 int max_sge;
186 unsigned int max_wrs;
187 unsigned int dma_local_lkey:1;
188 spinlock_t spinlock; /* protect the above */
189};
190
191/* bits for i_ack_flags */
192#define IB_ACK_IN_FLIGHT 0
193#define IB_ACK_REQUESTED 1
194
195/* Magic WR_ID for ACKs */
196#define RDS_IW_ACK_WR_ID ((u64)0xffffffffffffffffULL)
197#define RDS_IW_FAST_REG_WR_ID ((u64)0xefefefefefefefefULL)
198#define RDS_IW_LOCAL_INV_WR_ID ((u64)0xdfdfdfdfdfdfdfdfULL)
199
200struct rds_iw_statistics {
201 uint64_t s_iw_connect_raced;
202 uint64_t s_iw_listen_closed_stale;
203 uint64_t s_iw_tx_cq_call;
204 uint64_t s_iw_tx_cq_event;
205 uint64_t s_iw_tx_ring_full;
206 uint64_t s_iw_tx_throttle;
207 uint64_t s_iw_tx_sg_mapping_failure;
208 uint64_t s_iw_tx_stalled;
209 uint64_t s_iw_tx_credit_updates;
210 uint64_t s_iw_rx_cq_call;
211 uint64_t s_iw_rx_cq_event;
212 uint64_t s_iw_rx_ring_empty;
213 uint64_t s_iw_rx_refill_from_cq;
214 uint64_t s_iw_rx_refill_from_thread;
215 uint64_t s_iw_rx_alloc_limit;
216 uint64_t s_iw_rx_credit_updates;
217 uint64_t s_iw_ack_sent;
218 uint64_t s_iw_ack_send_failure;
219 uint64_t s_iw_ack_send_delayed;
220 uint64_t s_iw_ack_send_piggybacked;
221 uint64_t s_iw_ack_received;
222 uint64_t s_iw_rdma_mr_alloc;
223 uint64_t s_iw_rdma_mr_free;
224 uint64_t s_iw_rdma_mr_used;
225 uint64_t s_iw_rdma_mr_pool_flush;
226 uint64_t s_iw_rdma_mr_pool_wait;
227 uint64_t s_iw_rdma_mr_pool_depleted;
228};
229
230extern struct workqueue_struct *rds_iw_wq;
231
232/*
233 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
234 * doesn't define it.
235 */
236static inline void rds_iw_dma_sync_sg_for_cpu(struct ib_device *dev,
237 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
238{
239 unsigned int i;
240
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]),
245 direction);
246 }
247}
248#define ib_dma_sync_sg_for_cpu rds_iw_dma_sync_sg_for_cpu
249
250static inline void rds_iw_dma_sync_sg_for_device(struct ib_device *dev,
251 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
252{
253 unsigned int i;
254
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]),
259 direction);
260 }
261}
262#define ib_dma_sync_sg_for_device rds_iw_dma_sync_sg_for_device
263
264static inline u32 rds_iw_local_dma_lkey(struct rds_iw_connection *ic)
265{
266 return ic->i_dma_local_lkey ? ic->i_cm_id->device->local_dma_lkey : ic->i_mr->lkey;
267}
268
269/* ib.c */
270extern struct rds_transport rds_iw_transport;
271extern void rds_iw_add_one(struct ib_device *device);
272extern void rds_iw_remove_one(struct ib_device *device);
273extern struct ib_client rds_iw_client;
274
275extern unsigned int fastreg_pool_size;
276extern unsigned int fastreg_message_size;
277
278extern spinlock_t iw_nodev_conns_lock;
279extern struct list_head iw_nodev_conns;
280
281/* ib_cm.c */
282int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp);
283void rds_iw_conn_free(void *arg);
284int rds_iw_conn_connect(struct rds_connection *conn);
285void rds_iw_conn_shutdown(struct rds_connection *conn);
286void rds_iw_state_change(struct sock *sk);
287int __init rds_iw_listen_init(void);
288void rds_iw_listen_stop(void);
289void __rds_iw_conn_error(struct rds_connection *conn, const char *, ...);
290int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
291 struct rdma_cm_event *event);
292int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id);
293void rds_iw_cm_connect_complete(struct rds_connection *conn,
294 struct rdma_cm_event *event);
295
296
297#define rds_iw_conn_error(conn, fmt...) \
298 __rds_iw_conn_error(conn, KERN_WARNING "RDS/IW: " fmt)
299
300/* ib_rdma.c */
301int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
Andy Grover745cbcc2009-04-01 08:20:19 +0000302void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
303void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
304void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock);
305static inline void rds_iw_destroy_nodev_conns(void)
306{
307 __rds_iw_destroy_conns(&iw_nodev_conns, &iw_nodev_conns_lock);
308}
309static inline void rds_iw_destroy_conns(struct rds_iw_device *rds_iwdev)
310{
311 __rds_iw_destroy_conns(&rds_iwdev->conn_list, &rds_iwdev->spinlock);
312}
Andy Groverfcd8b7c2009-02-24 15:30:36 +0000313struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *);
314void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo);
315void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *);
316void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
317 struct rds_sock *rs, u32 *key_ret);
318void rds_iw_sync_mr(void *trans_private, int dir);
319void rds_iw_free_mr(void *trans_private, int invalidate);
320void rds_iw_flush_mrs(void);
321void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
322
323/* ib_recv.c */
324int __init rds_iw_recv_init(void);
325void rds_iw_recv_exit(void);
326int rds_iw_recv(struct rds_connection *conn);
327int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
328 gfp_t page_gfp, int prefill);
329void rds_iw_inc_purge(struct rds_incoming *inc);
330void rds_iw_inc_free(struct rds_incoming *inc);
331int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
332 size_t size);
333void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context);
334void rds_iw_recv_init_ring(struct rds_iw_connection *ic);
335void rds_iw_recv_clear_ring(struct rds_iw_connection *ic);
336void rds_iw_recv_init_ack(struct rds_iw_connection *ic);
337void rds_iw_attempt_ack(struct rds_iw_connection *ic);
338void rds_iw_ack_send_complete(struct rds_iw_connection *ic);
339u64 rds_iw_piggyb_ack(struct rds_iw_connection *ic);
340
341/* ib_ring.c */
342void rds_iw_ring_init(struct rds_iw_work_ring *ring, u32 nr);
343void rds_iw_ring_resize(struct rds_iw_work_ring *ring, u32 nr);
344u32 rds_iw_ring_alloc(struct rds_iw_work_ring *ring, u32 val, u32 *pos);
345void rds_iw_ring_free(struct rds_iw_work_ring *ring, u32 val);
346void rds_iw_ring_unalloc(struct rds_iw_work_ring *ring, u32 val);
347int rds_iw_ring_empty(struct rds_iw_work_ring *ring);
348int rds_iw_ring_low(struct rds_iw_work_ring *ring);
349u32 rds_iw_ring_oldest(struct rds_iw_work_ring *ring);
350u32 rds_iw_ring_completed(struct rds_iw_work_ring *ring, u32 wr_id, u32 oldest);
351extern wait_queue_head_t rds_iw_ring_empty_wait;
352
353/* ib_send.c */
354void rds_iw_xmit_complete(struct rds_connection *conn);
355int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
356 unsigned int hdr_off, unsigned int sg, unsigned int off);
357void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context);
358void rds_iw_send_init_ring(struct rds_iw_connection *ic);
359void rds_iw_send_clear_ring(struct rds_iw_connection *ic);
360int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op);
361void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits);
362void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted);
363int rds_iw_send_grab_credits(struct rds_iw_connection *ic, u32 wanted,
Steve Wise7b70d032009-04-09 14:09:39 +0000364 u32 *adv_credits, int need_posted, int max_posted);
Andy Groverfcd8b7c2009-02-24 15:30:36 +0000365
366/* ib_stats.c */
367DECLARE_PER_CPU(struct rds_iw_statistics, rds_iw_stats);
368#define rds_iw_stats_inc(member) rds_stats_inc_which(rds_iw_stats, member)
369unsigned int rds_iw_stats_info_copy(struct rds_info_iterator *iter,
370 unsigned int avail);
371
372/* ib_sysctl.c */
373int __init rds_iw_sysctl_init(void);
374void rds_iw_sysctl_exit(void);
375extern unsigned long rds_iw_sysctl_max_send_wr;
376extern unsigned long rds_iw_sysctl_max_recv_wr;
377extern unsigned long rds_iw_sysctl_max_unsig_wrs;
378extern unsigned long rds_iw_sysctl_max_unsig_bytes;
379extern unsigned long rds_iw_sysctl_max_recv_allocation;
380extern unsigned int rds_iw_sysctl_flow_control;
381extern ctl_table rds_iw_sysctl_table[];
382
383/*
384 * Helper functions for getting/setting the header and data SGEs in
385 * RDS packets (not RDMA)
386 */
387static inline struct ib_sge *
388rds_iw_header_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
389{
390 return &sge[0];
391}
392
393static inline struct ib_sge *
394rds_iw_data_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
395{
396 return &sge[1];
397}
398
Andy Groverfcd8b7c2009-02-24 15:30:36 +0000399#endif