1 /* SPDX-License-Identifier: GPL-2.0 */
5 #include <rdma/ib_verbs.h>
6 #include <rdma/rdma_cm.h>
7 #include <linux/interrupt.h>
9 #include <linux/slab.h>
11 #include "rdma_transport.h"
13 #define RDS_IB_MAX_SGE 8
14 #define RDS_IB_RECV_SGE 2
16 #define RDS_IB_DEFAULT_RECV_WR 1024
17 #define RDS_IB_DEFAULT_SEND_WR 256
18 #define RDS_IB_DEFAULT_FR_WR 256
19 #define RDS_IB_DEFAULT_FR_INV_WR 256
21 #define RDS_IB_DEFAULT_RETRY_COUNT 1
23 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
25 #define RDS_IB_RECYCLE_BATCH_COUNT 32
27 #define RDS_IB_WC_MAX 32
29 extern struct rw_semaphore rds_ib_devices_lock
;
30 extern struct list_head rds_ib_devices
;
33 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
34 * try and minimize the amount of memory tied up both the device and
35 * socket receive queues.
37 struct rds_page_frag
{
38 struct list_head f_item
;
39 struct list_head f_cache_entry
;
40 struct scatterlist f_sg
;
43 struct rds_ib_incoming
{
44 struct list_head ii_frags
;
45 struct list_head ii_cache_entry
;
46 struct rds_incoming ii_inc
;
49 struct rds_ib_cache_head
{
50 struct list_head
*first
;
54 struct rds_ib_refill_cache
{
55 struct rds_ib_cache_head __percpu
*percpu
;
56 struct list_head
*xfer
;
57 struct list_head
*ready
;
60 /* This is the common structure for the IB private data exchange in setting up
61 * an RDS connection. The exchange is different for IPv4 and IPv6 connections.
62 * The reason is that the address size is different and the addresses
63 * exchanged are in the beginning of the structure. Hence it is not possible
64 * for interoperability if same structure is used.
66 struct rds_ib_conn_priv_cmn
{
67 u8 ricpc_protocol_major
;
68 u8 ricpc_protocol_minor
;
69 __be16 ricpc_protocol_minor_mask
; /* bitmask */
70 __be32 ricpc_reserved1
;
72 __be32 ricpc_credit
; /* non-zero enables flow ctl */
75 struct rds_ib_connect_private
{
76 /* Add new fields at the end, and don't permute existing fields. */
79 struct rds_ib_conn_priv_cmn dp_cmn
;
82 struct rds6_ib_connect_private
{
83 /* Add new fields at the end, and don't permute existing fields. */
84 struct in6_addr dp_saddr
;
85 struct in6_addr dp_daddr
;
86 struct rds_ib_conn_priv_cmn dp_cmn
;
89 #define dp_protocol_major dp_cmn.ricpc_protocol_major
90 #define dp_protocol_minor dp_cmn.ricpc_protocol_minor
91 #define dp_protocol_minor_mask dp_cmn.ricpc_protocol_minor_mask
92 #define dp_ack_seq dp_cmn.ricpc_ack_seq
93 #define dp_credit dp_cmn.ricpc_credit
95 union rds_ib_conn_priv
{
96 struct rds_ib_connect_private ricp_v4
;
97 struct rds6_ib_connect_private ricp_v6
;
100 struct rds_ib_send_work
{
103 struct ib_send_wr s_wr
;
104 struct ib_rdma_wr s_rdma_wr
;
105 struct ib_atomic_wr s_atomic_wr
;
107 struct ib_sge s_sge
[RDS_IB_MAX_SGE
];
108 unsigned long s_queued
;
111 struct rds_ib_recv_work
{
112 struct rds_ib_incoming
*r_ibinc
;
113 struct rds_page_frag
*r_frag
;
114 struct ib_recv_wr r_wr
;
115 struct ib_sge r_sge
[2];
118 struct rds_ib_work_ring
{
126 /* Rings are posted with all the allocations they'll need to queue the
127 * incoming message to the receiving socket so this can't fail.
128 * All fragments start with a header, so we can make sure we're not receiving
129 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
131 struct rds_ib_ack_state
{
134 unsigned int ack_required
:1;
135 unsigned int ack_next_valid
:1;
136 unsigned int ack_recv_valid
:1;
140 struct rds_ib_device
;
142 struct rds_ib_connection
{
144 struct list_head ib_node
;
145 struct rds_ib_device
*rds_ibdev
;
146 struct rds_connection
*conn
;
148 /* alphabet soup, IBTA style */
149 struct rdma_cm_id
*i_cm_id
;
151 struct ib_cq
*i_send_cq
;
152 struct ib_cq
*i_recv_cq
;
153 struct ib_wc i_send_wc
[RDS_IB_WC_MAX
];
154 struct ib_wc i_recv_wc
[RDS_IB_WC_MAX
];
156 /* To control the number of wrs from fastreg */
157 atomic_t i_fastreg_wrs
;
158 atomic_t i_fastunreg_wrs
;
160 /* interrupt handling */
161 struct tasklet_struct i_send_tasklet
;
162 struct tasklet_struct i_recv_tasklet
;
165 struct rds_ib_work_ring i_send_ring
;
166 struct rm_data_op
*i_data_op
;
167 struct rds_header
*i_send_hdrs
;
168 dma_addr_t i_send_hdrs_dma
;
169 struct rds_ib_send_work
*i_sends
;
170 atomic_t i_signaled_sends
;
173 struct mutex i_recv_mutex
;
174 struct rds_ib_work_ring i_recv_ring
;
175 struct rds_ib_incoming
*i_ibinc
;
177 struct rds_header
*i_recv_hdrs
;
178 dma_addr_t i_recv_hdrs_dma
;
179 struct rds_ib_recv_work
*i_recvs
;
180 u64 i_ack_recv
; /* last ACK received */
181 struct rds_ib_refill_cache i_cache_incs
;
182 struct rds_ib_refill_cache i_cache_frags
;
183 atomic_t i_cache_allocs
;
186 unsigned long i_ack_flags
;
187 #ifdef KERNEL_HAS_ATOMIC64
188 atomic64_t i_ack_next
; /* next ACK to send */
190 spinlock_t i_ack_lock
; /* protect i_ack_next */
191 u64 i_ack_next
; /* next ACK to send */
193 struct rds_header
*i_ack
;
194 struct ib_send_wr i_ack_wr
;
195 struct ib_sge i_ack_sge
;
196 dma_addr_t i_ack_dma
;
197 unsigned long i_ack_queued
;
199 /* Flow control related information
201 * Our algorithm uses a pair variables that we need to access
202 * atomically - one for the send credits, and one posted
203 * recv credits we need to transfer to remote.
204 * Rather than protect them using a slow spinlock, we put both into
205 * a single atomic_t and update it using cmpxchg
209 /* Protocol version specific information */
210 unsigned int i_flowctl
:1; /* enable/disable flow ctl */
212 /* Batched completions */
213 unsigned int i_unsignaled_wrs
;
215 /* Endpoint role in connection */
217 atomic_t i_cq_quiesce
;
219 /* Send/Recv vectors */
224 /* This assumes that atomic_t is at least 32 bits */
225 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
226 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
227 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
228 #define IB_SET_POST_CREDITS(v) ((v) << 16)
230 struct rds_ib_ipaddr
{
231 struct list_head list
;
241 struct rds_ib_device
{
242 struct list_head list
;
243 struct list_head ipaddr_list
;
244 struct list_head conn_list
;
245 struct ib_device
*dev
;
249 unsigned int max_mrs
;
250 struct rds_ib_mr_pool
*mr_1m_pool
;
251 struct rds_ib_mr_pool
*mr_8k_pool
;
252 unsigned int fmr_max_remaps
;
253 unsigned int max_8k_mrs
;
254 unsigned int max_1m_mrs
;
256 unsigned int max_wrs
;
257 unsigned int max_initiator_depth
;
258 unsigned int max_responder_resources
;
259 spinlock_t spinlock
; /* protect the above */
261 struct work_struct free_work
;
265 #define ibdev_to_node(ibdev) dev_to_node((ibdev)->dev.parent)
266 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
268 /* bits for i_ack_flags */
269 #define IB_ACK_IN_FLIGHT 0
270 #define IB_ACK_REQUESTED 1
272 /* Magic WR_ID for ACKs */
273 #define RDS_IB_ACK_WR_ID (~(u64) 0)
275 struct rds_ib_statistics
{
276 uint64_t s_ib_connect_raced
;
277 uint64_t s_ib_listen_closed_stale
;
278 uint64_t s_ib_evt_handler_call
;
279 uint64_t s_ib_tasklet_call
;
280 uint64_t s_ib_tx_cq_event
;
281 uint64_t s_ib_tx_ring_full
;
282 uint64_t s_ib_tx_throttle
;
283 uint64_t s_ib_tx_sg_mapping_failure
;
284 uint64_t s_ib_tx_stalled
;
285 uint64_t s_ib_tx_credit_updates
;
286 uint64_t s_ib_rx_cq_event
;
287 uint64_t s_ib_rx_ring_empty
;
288 uint64_t s_ib_rx_refill_from_cq
;
289 uint64_t s_ib_rx_refill_from_thread
;
290 uint64_t s_ib_rx_alloc_limit
;
291 uint64_t s_ib_rx_total_frags
;
292 uint64_t s_ib_rx_total_incs
;
293 uint64_t s_ib_rx_credit_updates
;
294 uint64_t s_ib_ack_sent
;
295 uint64_t s_ib_ack_send_failure
;
296 uint64_t s_ib_ack_send_delayed
;
297 uint64_t s_ib_ack_send_piggybacked
;
298 uint64_t s_ib_ack_received
;
299 uint64_t s_ib_rdma_mr_8k_alloc
;
300 uint64_t s_ib_rdma_mr_8k_free
;
301 uint64_t s_ib_rdma_mr_8k_used
;
302 uint64_t s_ib_rdma_mr_8k_pool_flush
;
303 uint64_t s_ib_rdma_mr_8k_pool_wait
;
304 uint64_t s_ib_rdma_mr_8k_pool_depleted
;
305 uint64_t s_ib_rdma_mr_1m_alloc
;
306 uint64_t s_ib_rdma_mr_1m_free
;
307 uint64_t s_ib_rdma_mr_1m_used
;
308 uint64_t s_ib_rdma_mr_1m_pool_flush
;
309 uint64_t s_ib_rdma_mr_1m_pool_wait
;
310 uint64_t s_ib_rdma_mr_1m_pool_depleted
;
311 uint64_t s_ib_rdma_mr_8k_reused
;
312 uint64_t s_ib_rdma_mr_1m_reused
;
313 uint64_t s_ib_atomic_cswp
;
314 uint64_t s_ib_atomic_fadd
;
315 uint64_t s_ib_recv_added_to_cache
;
316 uint64_t s_ib_recv_removed_from_cache
;
319 extern struct workqueue_struct
*rds_ib_wq
;
322 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
325 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device
*dev
,
326 struct scatterlist
*sglist
,
327 unsigned int sg_dma_len
,
330 struct scatterlist
*sg
;
333 for_each_sg(sglist
, sg
, sg_dma_len
, i
) {
334 ib_dma_sync_single_for_cpu(dev
,
335 ib_sg_dma_address(dev
, sg
),
336 ib_sg_dma_len(dev
, sg
),
340 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
342 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device
*dev
,
343 struct scatterlist
*sglist
,
344 unsigned int sg_dma_len
,
347 struct scatterlist
*sg
;
350 for_each_sg(sglist
, sg
, sg_dma_len
, i
) {
351 ib_dma_sync_single_for_device(dev
,
352 ib_sg_dma_address(dev
, sg
),
353 ib_sg_dma_len(dev
, sg
),
357 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
361 extern struct rds_transport rds_ib_transport
;
362 struct rds_ib_device
*rds_ib_get_client_data(struct ib_device
*device
);
363 void rds_ib_dev_put(struct rds_ib_device
*rds_ibdev
);
364 extern struct ib_client rds_ib_client
;
366 extern unsigned int rds_ib_retry_count
;
368 extern spinlock_t ib_nodev_conns_lock
;
369 extern struct list_head ib_nodev_conns
;
372 int rds_ib_conn_alloc(struct rds_connection
*conn
, gfp_t gfp
);
373 void rds_ib_conn_free(void *arg
);
374 int rds_ib_conn_path_connect(struct rds_conn_path
*cp
);
375 void rds_ib_conn_path_shutdown(struct rds_conn_path
*cp
);
376 void rds_ib_state_change(struct sock
*sk
);
377 int rds_ib_listen_init(void);
378 void rds_ib_listen_stop(void);
380 void __rds_ib_conn_error(struct rds_connection
*conn
, const char *, ...);
381 int rds_ib_cm_handle_connect(struct rdma_cm_id
*cm_id
,
382 struct rdma_cm_event
*event
, bool isv6
);
383 int rds_ib_cm_initiate_connect(struct rdma_cm_id
*cm_id
, bool isv6
);
384 void rds_ib_cm_connect_complete(struct rds_connection
*conn
,
385 struct rdma_cm_event
*event
);
388 #define rds_ib_conn_error(conn, fmt...) \
389 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
392 int rds_ib_update_ipaddr(struct rds_ib_device
*rds_ibdev
,
393 struct in6_addr
*ipaddr
);
394 void rds_ib_add_conn(struct rds_ib_device
*rds_ibdev
, struct rds_connection
*conn
);
395 void rds_ib_remove_conn(struct rds_ib_device
*rds_ibdev
, struct rds_connection
*conn
);
396 void rds_ib_destroy_nodev_conns(void);
397 void rds_ib_mr_cqe_handler(struct rds_ib_connection
*ic
, struct ib_wc
*wc
);
400 int rds_ib_recv_init(void);
401 void rds_ib_recv_exit(void);
402 int rds_ib_recv_path(struct rds_conn_path
*conn
);
403 int rds_ib_recv_alloc_caches(struct rds_ib_connection
*ic
, gfp_t gfp
);
404 void rds_ib_recv_free_caches(struct rds_ib_connection
*ic
);
405 void rds_ib_recv_refill(struct rds_connection
*conn
, int prefill
, gfp_t gfp
);
406 void rds_ib_inc_free(struct rds_incoming
*inc
);
407 int rds_ib_inc_copy_to_user(struct rds_incoming
*inc
, struct iov_iter
*to
);
408 void rds_ib_recv_cqe_handler(struct rds_ib_connection
*ic
, struct ib_wc
*wc
,
409 struct rds_ib_ack_state
*state
);
410 void rds_ib_recv_tasklet_fn(unsigned long data
);
411 void rds_ib_recv_init_ring(struct rds_ib_connection
*ic
);
412 void rds_ib_recv_clear_ring(struct rds_ib_connection
*ic
);
413 void rds_ib_recv_init_ack(struct rds_ib_connection
*ic
);
414 void rds_ib_attempt_ack(struct rds_ib_connection
*ic
);
415 void rds_ib_ack_send_complete(struct rds_ib_connection
*ic
);
416 u64
rds_ib_piggyb_ack(struct rds_ib_connection
*ic
);
417 void rds_ib_set_ack(struct rds_ib_connection
*ic
, u64 seq
, int ack_required
);
420 void rds_ib_ring_init(struct rds_ib_work_ring
*ring
, u32 nr
);
421 void rds_ib_ring_resize(struct rds_ib_work_ring
*ring
, u32 nr
);
422 u32
rds_ib_ring_alloc(struct rds_ib_work_ring
*ring
, u32 val
, u32
*pos
);
423 void rds_ib_ring_free(struct rds_ib_work_ring
*ring
, u32 val
);
424 void rds_ib_ring_unalloc(struct rds_ib_work_ring
*ring
, u32 val
);
425 int rds_ib_ring_empty(struct rds_ib_work_ring
*ring
);
426 int rds_ib_ring_low(struct rds_ib_work_ring
*ring
);
427 u32
rds_ib_ring_oldest(struct rds_ib_work_ring
*ring
);
428 u32
rds_ib_ring_completed(struct rds_ib_work_ring
*ring
, u32 wr_id
, u32 oldest
);
429 extern wait_queue_head_t rds_ib_ring_empty_wait
;
432 void rds_ib_xmit_path_complete(struct rds_conn_path
*cp
);
433 int rds_ib_xmit(struct rds_connection
*conn
, struct rds_message
*rm
,
434 unsigned int hdr_off
, unsigned int sg
, unsigned int off
);
435 void rds_ib_send_cqe_handler(struct rds_ib_connection
*ic
, struct ib_wc
*wc
);
436 void rds_ib_send_init_ring(struct rds_ib_connection
*ic
);
437 void rds_ib_send_clear_ring(struct rds_ib_connection
*ic
);
438 int rds_ib_xmit_rdma(struct rds_connection
*conn
, struct rm_rdma_op
*op
);
439 void rds_ib_send_add_credits(struct rds_connection
*conn
, unsigned int credits
);
440 void rds_ib_advertise_credits(struct rds_connection
*conn
, unsigned int posted
);
441 int rds_ib_send_grab_credits(struct rds_ib_connection
*ic
, u32 wanted
,
442 u32
*adv_credits
, int need_posted
, int max_posted
);
443 int rds_ib_xmit_atomic(struct rds_connection
*conn
, struct rm_atomic_op
*op
);
446 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_ib_statistics
, rds_ib_stats
);
447 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
448 #define rds_ib_stats_add(member, count) \
449 rds_stats_add_which(rds_ib_stats, member, count)
450 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator
*iter
,
454 int rds_ib_sysctl_init(void);
455 void rds_ib_sysctl_exit(void);
456 extern unsigned long rds_ib_sysctl_max_send_wr
;
457 extern unsigned long rds_ib_sysctl_max_recv_wr
;
458 extern unsigned long rds_ib_sysctl_max_unsig_wrs
;
459 extern unsigned long rds_ib_sysctl_max_unsig_bytes
;
460 extern unsigned long rds_ib_sysctl_max_recv_allocation
;
461 extern unsigned int rds_ib_sysctl_flow_control
;