tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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 <linux/interrupt.h>
7#include <linux/pci.h>
8#include <linux/slab.h>
9#include "rds.h"
10#include "rdma_transport.h"
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#define RDS_IB_DEFAULT_FR_WR 512
18
19#define RDS_IB_DEFAULT_RETRY_COUNT 2
20
21#define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
22
23#define RDS_IB_RECYCLE_BATCH_COUNT 32
24
25#define RDS_IB_WC_MAX 32
26
27extern struct rw_semaphore rds_ib_devices_lock;
28extern struct list_head rds_ib_devices;
29
30/*
31 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
32 * try and minimize the amount of memory tied up both the device and
33 * socket receive queues.
34 */
35struct rds_page_frag {
36 struct list_head f_item;
37 struct list_head f_cache_entry;
38 struct scatterlist f_sg;
39};
40
41struct rds_ib_incoming {
42 struct list_head ii_frags;
43 struct list_head ii_cache_entry;
44 struct rds_incoming ii_inc;
45};
46
47struct rds_ib_cache_head {
48 struct list_head *first;
49 unsigned long count;
50};
51
52struct rds_ib_refill_cache {
53 struct rds_ib_cache_head __percpu *percpu;
54 struct list_head *xfer;
55 struct list_head *ready;
56};
57
58struct rds_ib_connect_private {
59 /* Add new fields at the end, and don't permute existing fields. */
60 __be32 dp_saddr;
61 __be32 dp_daddr;
62 u8 dp_protocol_major;
63 u8 dp_protocol_minor;
64 __be16 dp_protocol_minor_mask; /* bitmask */
65 __be32 dp_reserved1;
66 __be64 dp_ack_seq;
67 __be32 dp_credit; /* non-zero enables flow ctl */
68};
69
70struct rds_ib_send_work {
71 void *s_op;
72 union {
73 struct ib_send_wr s_wr;
74 struct ib_rdma_wr s_rdma_wr;
75 struct ib_atomic_wr s_atomic_wr;
76 };
77 struct ib_sge s_sge[RDS_IB_MAX_SGE];
78 unsigned long s_queued;
79};
80
81struct rds_ib_recv_work {
82 struct rds_ib_incoming *r_ibinc;
83 struct rds_page_frag *r_frag;
84 struct ib_recv_wr r_wr;
85 struct ib_sge r_sge[2];
86};
87
88struct rds_ib_work_ring {
89 u32 w_nr;
90 u32 w_alloc_ptr;
91 u32 w_alloc_ctr;
92 u32 w_free_ptr;
93 atomic_t w_free_ctr;
94};
95
96/* Rings are posted with all the allocations they'll need to queue the
97 * incoming message to the receiving socket so this can't fail.
98 * All fragments start with a header, so we can make sure we're not receiving
99 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
100 */
101struct rds_ib_ack_state {
102 u64 ack_next;
103 u64 ack_recv;
104 unsigned int ack_required:1;
105 unsigned int ack_next_valid:1;
106 unsigned int ack_recv_valid:1;
107};
108
109
110struct rds_ib_device;
111
112struct rds_ib_connection {
113
114 struct list_head ib_node;
115 struct rds_ib_device *rds_ibdev;
116 struct rds_connection *conn;
117
118 /* alphabet soup, IBTA style */
119 struct rdma_cm_id *i_cm_id;
120 struct ib_pd *i_pd;
121 struct ib_cq *i_send_cq;
122 struct ib_cq *i_recv_cq;
123 struct ib_wc i_send_wc[RDS_IB_WC_MAX];
124 struct ib_wc i_recv_wc[RDS_IB_WC_MAX];
125
126 /* To control the number of wrs from fastreg */
127 atomic_t i_fastreg_wrs;
128
129 /* interrupt handling */
130 struct tasklet_struct i_send_tasklet;
131 struct tasklet_struct i_recv_tasklet;
132
133 /* tx */
134 struct rds_ib_work_ring i_send_ring;
135 struct rm_data_op *i_data_op;
136 struct rds_header *i_send_hdrs;
137 u64 i_send_hdrs_dma;
138 struct rds_ib_send_work *i_sends;
139 atomic_t i_signaled_sends;
140
141 /* rx */
142 struct mutex i_recv_mutex;
143 struct rds_ib_work_ring i_recv_ring;
144 struct rds_ib_incoming *i_ibinc;
145 u32 i_recv_data_rem;
146 struct rds_header *i_recv_hdrs;
147 u64 i_recv_hdrs_dma;
148 struct rds_ib_recv_work *i_recvs;
149 u64 i_ack_recv; /* last ACK received */
150 struct rds_ib_refill_cache i_cache_incs;
151 struct rds_ib_refill_cache i_cache_frags;
152
153 /* sending acks */
154 unsigned long i_ack_flags;
155#ifdef KERNEL_HAS_ATOMIC64
156 atomic64_t i_ack_next; /* next ACK to send */
157#else
158 spinlock_t i_ack_lock; /* protect i_ack_next */
159 u64 i_ack_next; /* next ACK to send */
160#endif
161 struct rds_header *i_ack;
162 struct ib_send_wr i_ack_wr;
163 struct ib_sge i_ack_sge;
164 u64 i_ack_dma;
165 unsigned long i_ack_queued;
166
167 /* Flow control related information
168 *
169 * Our algorithm uses a pair variables that we need to access
170 * atomically - one for the send credits, and one posted
171 * recv credits we need to transfer to remote.
172 * Rather than protect them using a slow spinlock, we put both into
173 * a single atomic_t and update it using cmpxchg
174 */
175 atomic_t i_credits;
176
177 /* Protocol version specific information */
178 unsigned int i_flowctl:1; /* enable/disable flow ctl */
179
180 /* Batched completions */
181 unsigned int i_unsignaled_wrs;
182};
183
184/* This assumes that atomic_t is at least 32 bits */
185#define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
186#define IB_GET_POST_CREDITS(v) ((v) >> 16)
187#define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
188#define IB_SET_POST_CREDITS(v) ((v) << 16)
189
190struct rds_ib_ipaddr {
191 struct list_head list;
192 __be32 ipaddr;
193 struct rcu_head rcu;
194};
195
196enum {
197 RDS_IB_MR_8K_POOL,
198 RDS_IB_MR_1M_POOL,
199};
200
201struct rds_ib_device {
202 struct list_head list;
203 struct list_head ipaddr_list;
204 struct list_head conn_list;
205 struct ib_device *dev;
206 struct ib_pd *pd;
207 bool has_fmr;
208 bool has_fr;
209 bool use_fastreg;
210
211 unsigned int max_mrs;
212 struct rds_ib_mr_pool *mr_1m_pool;
213 struct rds_ib_mr_pool *mr_8k_pool;
214 unsigned int fmr_max_remaps;
215 unsigned int max_8k_mrs;
216 unsigned int max_1m_mrs;
217 int max_sge;
218 unsigned int max_wrs;
219 unsigned int max_initiator_depth;
220 unsigned int max_responder_resources;
221 spinlock_t spinlock; /* protect the above */
222 atomic_t refcount;
223 struct work_struct free_work;
224};
225
226#define ibdev_to_node(ibdev) dev_to_node(ibdev->dma_device)
227#define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
228
229/* bits for i_ack_flags */
230#define IB_ACK_IN_FLIGHT 0
231#define IB_ACK_REQUESTED 1
232
233/* Magic WR_ID for ACKs */
234#define RDS_IB_ACK_WR_ID (~(u64) 0)
235
236struct rds_ib_statistics {
237 uint64_t s_ib_connect_raced;
238 uint64_t s_ib_listen_closed_stale;
239 uint64_t s_ib_evt_handler_call;
240 uint64_t s_ib_tasklet_call;
241 uint64_t s_ib_tx_cq_event;
242 uint64_t s_ib_tx_ring_full;
243 uint64_t s_ib_tx_throttle;
244 uint64_t s_ib_tx_sg_mapping_failure;
245 uint64_t s_ib_tx_stalled;
246 uint64_t s_ib_tx_credit_updates;
247 uint64_t s_ib_rx_cq_event;
248 uint64_t s_ib_rx_ring_empty;
249 uint64_t s_ib_rx_refill_from_cq;
250 uint64_t s_ib_rx_refill_from_thread;
251 uint64_t s_ib_rx_alloc_limit;
252 uint64_t s_ib_rx_credit_updates;
253 uint64_t s_ib_ack_sent;
254 uint64_t s_ib_ack_send_failure;
255 uint64_t s_ib_ack_send_delayed;
256 uint64_t s_ib_ack_send_piggybacked;
257 uint64_t s_ib_ack_received;
258 uint64_t s_ib_rdma_mr_8k_alloc;
259 uint64_t s_ib_rdma_mr_8k_free;
260 uint64_t s_ib_rdma_mr_8k_used;
261 uint64_t s_ib_rdma_mr_8k_pool_flush;
262 uint64_t s_ib_rdma_mr_8k_pool_wait;
263 uint64_t s_ib_rdma_mr_8k_pool_depleted;
264 uint64_t s_ib_rdma_mr_1m_alloc;
265 uint64_t s_ib_rdma_mr_1m_free;
266 uint64_t s_ib_rdma_mr_1m_used;
267 uint64_t s_ib_rdma_mr_1m_pool_flush;
268 uint64_t s_ib_rdma_mr_1m_pool_wait;
269 uint64_t s_ib_rdma_mr_1m_pool_depleted;
270 uint64_t s_ib_rdma_mr_8k_reused;
271 uint64_t s_ib_rdma_mr_1m_reused;
272 uint64_t s_ib_atomic_cswp;
273 uint64_t s_ib_atomic_fadd;
274};
275
276extern struct workqueue_struct *rds_ib_wq;
277
278/*
279 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
280 * doesn't define it.
281 */
282static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
283 struct scatterlist *sglist,
284 unsigned int sg_dma_len,
285 int direction)
286{
287 struct scatterlist *sg;
288 unsigned int i;
289
290 for_each_sg(sglist, sg, sg_dma_len, i) {
291 ib_dma_sync_single_for_cpu(dev,
292 ib_sg_dma_address(dev, sg),
293 ib_sg_dma_len(dev, sg),
294 direction);
295 }
296}
297#define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
298
299static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
300 struct scatterlist *sglist,
301 unsigned int sg_dma_len,
302 int direction)
303{
304 struct scatterlist *sg;
305 unsigned int i;
306
307 for_each_sg(sglist, sg, sg_dma_len, i) {
308 ib_dma_sync_single_for_device(dev,
309 ib_sg_dma_address(dev, sg),
310 ib_sg_dma_len(dev, sg),
311 direction);
312 }
313}
314#define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
315
316
317/* ib.c */
318extern struct rds_transport rds_ib_transport;
319struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
320void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
321extern struct ib_client rds_ib_client;
322
323extern unsigned int rds_ib_retry_count;
324
325extern spinlock_t ib_nodev_conns_lock;
326extern struct list_head ib_nodev_conns;
327
328/* ib_cm.c */
329int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
330void rds_ib_conn_free(void *arg);
331int rds_ib_conn_path_connect(struct rds_conn_path *cp);
332void rds_ib_conn_path_shutdown(struct rds_conn_path *cp);
333void rds_ib_state_change(struct sock *sk);
334int rds_ib_listen_init(void);
335void rds_ib_listen_stop(void);
336void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
337int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
338 struct rdma_cm_event *event);
339int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
340void rds_ib_cm_connect_complete(struct rds_connection *conn,
341 struct rdma_cm_event *event);
342
343
344#define rds_ib_conn_error(conn, fmt...) \
345 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
346
347/* ib_rdma.c */
348int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
349void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
350void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
351void rds_ib_destroy_nodev_conns(void);
352void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
353
354/* ib_recv.c */
355int rds_ib_recv_init(void);
356void rds_ib_recv_exit(void);
357int rds_ib_recv_path(struct rds_conn_path *conn);
358int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic);
359void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
360void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp);
361void rds_ib_inc_free(struct rds_incoming *inc);
362int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
363void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc,
364 struct rds_ib_ack_state *state);
365void rds_ib_recv_tasklet_fn(unsigned long data);
366void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
367void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
368void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
369void rds_ib_attempt_ack(struct rds_ib_connection *ic);
370void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
371u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
372void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required);
373
374/* ib_ring.c */
375void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
376void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
377u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
378void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
379void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
380int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
381int rds_ib_ring_low(struct rds_ib_work_ring *ring);
382u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
383u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
384extern wait_queue_head_t rds_ib_ring_empty_wait;
385
386/* ib_send.c */
387void rds_ib_xmit_path_complete(struct rds_conn_path *cp);
388int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
389 unsigned int hdr_off, unsigned int sg, unsigned int off);
390void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
391void rds_ib_send_init_ring(struct rds_ib_connection *ic);
392void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
393int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
394void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
395void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
396int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
397 u32 *adv_credits, int need_posted, int max_posted);
398int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
399
400/* ib_stats.c */
401DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
402#define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
403unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
404 unsigned int avail);
405
406/* ib_sysctl.c */
407int rds_ib_sysctl_init(void);
408void rds_ib_sysctl_exit(void);
409extern unsigned long rds_ib_sysctl_max_send_wr;
410extern unsigned long rds_ib_sysctl_max_recv_wr;
411extern unsigned long rds_ib_sysctl_max_unsig_wrs;
412extern unsigned long rds_ib_sysctl_max_unsig_bytes;
413extern unsigned long rds_ib_sysctl_max_recv_allocation;
414extern unsigned int rds_ib_sysctl_flow_control;
415
416#endif