4 #include <rdma/ib_verbs.h>
5 #include <rdma/rdma_cm.h>
7 #include <linux/slab.h>
9 #include "rdma_transport.h"
11 #define RDS_FMR_SIZE 256
12 #define RDS_FMR_POOL_SIZE 8192
14 #define RDS_IB_MAX_SGE 8
15 #define RDS_IB_RECV_SGE 2
17 #define RDS_IB_DEFAULT_RECV_WR 1024
18 #define RDS_IB_DEFAULT_SEND_WR 256
20 #define RDS_IB_DEFAULT_RETRY_COUNT 2
22 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
24 #define RDS_IB_RECYCLE_BATCH_COUNT 32
26 extern struct list_head rds_ib_devices
;
29 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
30 * try and minimize the amount of memory tied up both the device and
31 * socket receive queues.
33 struct rds_page_frag
{
34 struct list_head f_item
;
35 struct list_head f_cache_entry
;
36 struct scatterlist f_sg
;
39 struct rds_ib_incoming
{
40 struct list_head ii_frags
;
41 struct list_head ii_cache_entry
;
42 struct rds_incoming ii_inc
;
45 struct rds_ib_cache_head
{
46 struct list_head
*first
;
50 struct rds_ib_refill_cache
{
51 struct rds_ib_cache_head
*percpu
;
52 struct list_head
*xfer
;
53 struct list_head
*ready
;
56 struct rds_ib_connect_private
{
57 /* Add new fields at the end, and don't permute existing fields. */
62 __be16 dp_protocol_minor_mask
; /* bitmask */
65 __be32 dp_credit
; /* non-zero enables flow ctl */
68 struct rds_ib_send_work
{
70 struct ib_send_wr s_wr
;
71 struct ib_sge s_sge
[RDS_IB_MAX_SGE
];
72 unsigned long s_queued
;
75 struct rds_ib_recv_work
{
76 struct rds_ib_incoming
*r_ibinc
;
77 struct rds_page_frag
*r_frag
;
78 struct ib_recv_wr r_wr
;
79 struct ib_sge r_sge
[2];
82 struct rds_ib_work_ring
{
92 struct rds_ib_connection
{
94 struct list_head ib_node
;
95 struct rds_ib_device
*rds_ibdev
;
96 struct rds_connection
*conn
;
98 /* alphabet soup, IBTA style */
99 struct rdma_cm_id
*i_cm_id
;
102 struct ib_cq
*i_send_cq
;
103 struct ib_cq
*i_recv_cq
;
106 struct rds_ib_work_ring i_send_ring
;
107 struct rm_data_op
*i_data_op
;
108 struct rds_header
*i_send_hdrs
;
110 struct rds_ib_send_work
*i_sends
;
113 struct tasklet_struct i_recv_tasklet
;
114 struct mutex i_recv_mutex
;
115 struct rds_ib_work_ring i_recv_ring
;
116 struct rds_ib_incoming
*i_ibinc
;
118 struct rds_header
*i_recv_hdrs
;
120 struct rds_ib_recv_work
*i_recvs
;
121 u64 i_ack_recv
; /* last ACK received */
122 struct rds_ib_refill_cache i_cache_incs
;
123 struct rds_ib_refill_cache i_cache_frags
;
126 unsigned long i_ack_flags
;
127 #ifdef KERNEL_HAS_ATOMIC64
128 atomic64_t i_ack_next
; /* next ACK to send */
130 spinlock_t i_ack_lock
; /* protect i_ack_next */
131 u64 i_ack_next
; /* next ACK to send */
133 struct rds_header
*i_ack
;
134 struct ib_send_wr i_ack_wr
;
135 struct ib_sge i_ack_sge
;
137 unsigned long i_ack_queued
;
139 /* Flow control related information
141 * Our algorithm uses a pair variables that we need to access
142 * atomically - one for the send credits, and one posted
143 * recv credits we need to transfer to remote.
144 * Rather than protect them using a slow spinlock, we put both into
145 * a single atomic_t and update it using cmpxchg
149 /* Protocol version specific information */
150 unsigned int i_flowctl
:1; /* enable/disable flow ctl */
152 /* Batched completions */
153 unsigned int i_unsignaled_wrs
;
156 /* This assumes that atomic_t is at least 32 bits */
157 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
158 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
159 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
160 #define IB_SET_POST_CREDITS(v) ((v) << 16)
162 struct rds_ib_ipaddr
{
163 struct list_head list
;
167 struct rds_ib_device
{
168 struct list_head list
;
169 struct list_head ipaddr_list
;
170 struct list_head conn_list
;
171 struct ib_device
*dev
;
174 struct rds_ib_mr_pool
*mr_pool
;
175 unsigned int fmr_max_remaps
;
176 unsigned int max_fmrs
;
178 unsigned int max_wrs
;
179 unsigned int max_initiator_depth
;
180 unsigned int max_responder_resources
;
181 spinlock_t spinlock
; /* protect the above */
183 struct work_struct free_work
;
186 #define pcidev_to_node(pcidev) pcibus_to_node(pcidev->bus)
187 #define ibdev_to_node(ibdev) pcidev_to_node(to_pci_dev(ibdev->dma_device))
188 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
190 /* bits for i_ack_flags */
191 #define IB_ACK_IN_FLIGHT 0
192 #define IB_ACK_REQUESTED 1
194 /* Magic WR_ID for ACKs */
195 #define RDS_IB_ACK_WR_ID (~(u64) 0)
197 struct rds_ib_statistics
{
198 uint64_t s_ib_connect_raced
;
199 uint64_t s_ib_listen_closed_stale
;
200 uint64_t s_ib_tx_cq_call
;
201 uint64_t s_ib_tx_cq_event
;
202 uint64_t s_ib_tx_ring_full
;
203 uint64_t s_ib_tx_throttle
;
204 uint64_t s_ib_tx_sg_mapping_failure
;
205 uint64_t s_ib_tx_stalled
;
206 uint64_t s_ib_tx_credit_updates
;
207 uint64_t s_ib_rx_cq_call
;
208 uint64_t s_ib_rx_cq_event
;
209 uint64_t s_ib_rx_ring_empty
;
210 uint64_t s_ib_rx_refill_from_cq
;
211 uint64_t s_ib_rx_refill_from_thread
;
212 uint64_t s_ib_rx_alloc_limit
;
213 uint64_t s_ib_rx_credit_updates
;
214 uint64_t s_ib_ack_sent
;
215 uint64_t s_ib_ack_send_failure
;
216 uint64_t s_ib_ack_send_delayed
;
217 uint64_t s_ib_ack_send_piggybacked
;
218 uint64_t s_ib_ack_received
;
219 uint64_t s_ib_rdma_mr_alloc
;
220 uint64_t s_ib_rdma_mr_free
;
221 uint64_t s_ib_rdma_mr_used
;
222 uint64_t s_ib_rdma_mr_pool_flush
;
223 uint64_t s_ib_rdma_mr_pool_wait
;
224 uint64_t s_ib_rdma_mr_pool_depleted
;
225 uint64_t s_ib_atomic_cswp
;
226 uint64_t s_ib_atomic_fadd
;
229 extern struct workqueue_struct
*rds_ib_wq
;
232 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
235 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device
*dev
,
236 struct scatterlist
*sg
, unsigned int sg_dma_len
, int direction
)
240 for (i
= 0; i
< sg_dma_len
; ++i
) {
241 ib_dma_sync_single_for_cpu(dev
,
242 ib_sg_dma_address(dev
, &sg
[i
]),
243 ib_sg_dma_len(dev
, &sg
[i
]),
247 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
249 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device
*dev
,
250 struct scatterlist
*sg
, unsigned int sg_dma_len
, int direction
)
254 for (i
= 0; i
< sg_dma_len
; ++i
) {
255 ib_dma_sync_single_for_device(dev
,
256 ib_sg_dma_address(dev
, &sg
[i
]),
257 ib_sg_dma_len(dev
, &sg
[i
]),
261 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
265 extern struct rds_transport rds_ib_transport
;
266 extern void rds_ib_add_one(struct ib_device
*device
);
267 extern void rds_ib_remove_one(struct ib_device
*device
);
268 struct rds_ib_device
*rds_ib_get_client_data(struct ib_device
*device
);
269 void rds_ib_dev_put(struct rds_ib_device
*rds_ibdev
);
270 extern struct ib_client rds_ib_client
;
272 extern unsigned int fmr_pool_size
;
273 extern unsigned int fmr_message_size
;
274 extern unsigned int rds_ib_retry_count
;
276 extern spinlock_t ib_nodev_conns_lock
;
277 extern struct list_head ib_nodev_conns
;
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 __init
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
);
295 #define rds_ib_conn_error(conn, fmt...) \
296 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
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_conns(struct list_head
*list
, spinlock_t
*list_lock
);
303 static inline void rds_ib_destroy_nodev_conns(void)
305 __rds_ib_destroy_conns(&ib_nodev_conns
, &ib_nodev_conns_lock
);
307 static inline void rds_ib_destroy_conns(struct rds_ib_device
*rds_ibdev
)
309 __rds_ib_destroy_conns(&rds_ibdev
->conn_list
, &rds_ibdev
->spinlock
);
311 struct rds_ib_mr_pool
*rds_ib_create_mr_pool(struct rds_ib_device
*);
312 void rds_ib_get_mr_info(struct rds_ib_device
*rds_ibdev
, struct rds_info_rdma_connection
*iinfo
);
313 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool
*);
314 void *rds_ib_get_mr(struct scatterlist
*sg
, unsigned long nents
,
315 struct rds_sock
*rs
, u32
*key_ret
);
316 void rds_ib_sync_mr(void *trans_private
, int dir
);
317 void rds_ib_free_mr(void *trans_private
, int invalidate
);
318 void rds_ib_flush_mrs(void);
321 int __init
rds_ib_recv_init(void);
322 void rds_ib_recv_exit(void);
323 int rds_ib_recv(struct rds_connection
*conn
);
324 int rds_ib_recv_alloc_caches(struct rds_ib_connection
*ic
);
325 void rds_ib_recv_free_caches(struct rds_ib_connection
*ic
);
326 int rds_ib_recv_refill(struct rds_connection
*conn
, int prefill
);
327 void rds_ib_inc_free(struct rds_incoming
*inc
);
328 int rds_ib_inc_copy_to_user(struct rds_incoming
*inc
, struct iovec
*iov
,
330 void rds_ib_recv_cq_comp_handler(struct ib_cq
*cq
, void *context
);
331 void rds_ib_recv_tasklet_fn(unsigned long data
);
332 void rds_ib_recv_init_ring(struct rds_ib_connection
*ic
);
333 void rds_ib_recv_clear_ring(struct rds_ib_connection
*ic
);
334 void rds_ib_recv_init_ack(struct rds_ib_connection
*ic
);
335 void rds_ib_attempt_ack(struct rds_ib_connection
*ic
);
336 void rds_ib_ack_send_complete(struct rds_ib_connection
*ic
);
337 u64
rds_ib_piggyb_ack(struct rds_ib_connection
*ic
);
340 void rds_ib_ring_init(struct rds_ib_work_ring
*ring
, u32 nr
);
341 void rds_ib_ring_resize(struct rds_ib_work_ring
*ring
, u32 nr
);
342 u32
rds_ib_ring_alloc(struct rds_ib_work_ring
*ring
, u32 val
, u32
*pos
);
343 void rds_ib_ring_free(struct rds_ib_work_ring
*ring
, u32 val
);
344 void rds_ib_ring_unalloc(struct rds_ib_work_ring
*ring
, u32 val
);
345 int rds_ib_ring_empty(struct rds_ib_work_ring
*ring
);
346 int rds_ib_ring_low(struct rds_ib_work_ring
*ring
);
347 u32
rds_ib_ring_oldest(struct rds_ib_work_ring
*ring
);
348 u32
rds_ib_ring_completed(struct rds_ib_work_ring
*ring
, u32 wr_id
, u32 oldest
);
349 extern wait_queue_head_t rds_ib_ring_empty_wait
;
352 void rds_ib_xmit_complete(struct rds_connection
*conn
);
353 int rds_ib_xmit(struct rds_connection
*conn
, struct rds_message
*rm
,
354 unsigned int hdr_off
, unsigned int sg
, unsigned int off
);
355 void rds_ib_send_cq_comp_handler(struct ib_cq
*cq
, void *context
);
356 void rds_ib_send_init_ring(struct rds_ib_connection
*ic
);
357 void rds_ib_send_clear_ring(struct rds_ib_connection
*ic
);
358 int rds_ib_xmit_rdma(struct rds_connection
*conn
, struct rm_rdma_op
*op
);
359 void rds_ib_send_add_credits(struct rds_connection
*conn
, unsigned int credits
);
360 void rds_ib_advertise_credits(struct rds_connection
*conn
, unsigned int posted
);
361 int rds_ib_send_grab_credits(struct rds_ib_connection
*ic
, u32 wanted
,
362 u32
*adv_credits
, int need_posted
, int max_posted
);
363 int rds_ib_xmit_atomic(struct rds_connection
*conn
, struct rm_atomic_op
*op
);
366 DECLARE_PER_CPU(struct rds_ib_statistics
, rds_ib_stats
);
367 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
368 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator
*iter
,
372 int __init
rds_ib_sysctl_init(void);
373 void rds_ib_sysctl_exit(void);
374 extern unsigned long rds_ib_sysctl_max_send_wr
;
375 extern unsigned long rds_ib_sysctl_max_recv_wr
;
376 extern unsigned long rds_ib_sysctl_max_unsig_wrs
;
377 extern unsigned long rds_ib_sysctl_max_unsig_bytes
;
378 extern unsigned long rds_ib_sysctl_max_recv_allocation
;
379 extern unsigned int rds_ib_sysctl_flow_control
;
380 extern ctl_table rds_ib_sysctl_table
[];