net-gro: reset skb->pkt_type in napi_reuse_skb()
[linux-stable.git] / net / rds / ib.h
blob7db93f7f5c616a5e176500cd0d13f784a88b3ed8
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _RDS_IB_H
3 #define _RDS_IB_H
5 #include <rdma/ib_verbs.h>
6 #include <rdma/rdma_cm.h>
7 #include <linux/interrupt.h>
8 #include <linux/pci.h>
9 #include <linux/slab.h>
10 #include "rds.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;
51 unsigned long count;
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 struct rds_ib_connect_private {
61 /* Add new fields at the end, and don't permute existing fields. */
62 __be32 dp_saddr;
63 __be32 dp_daddr;
64 u8 dp_protocol_major;
65 u8 dp_protocol_minor;
66 __be16 dp_protocol_minor_mask; /* bitmask */
67 __be32 dp_reserved1;
68 __be64 dp_ack_seq;
69 __be32 dp_credit; /* non-zero enables flow ctl */
72 struct rds_ib_send_work {
73 void *s_op;
74 union {
75 struct ib_send_wr s_wr;
76 struct ib_rdma_wr s_rdma_wr;
77 struct ib_atomic_wr s_atomic_wr;
79 struct ib_sge s_sge[RDS_IB_MAX_SGE];
80 unsigned long s_queued;
83 struct rds_ib_recv_work {
84 struct rds_ib_incoming *r_ibinc;
85 struct rds_page_frag *r_frag;
86 struct ib_recv_wr r_wr;
87 struct ib_sge r_sge[2];
90 struct rds_ib_work_ring {
91 u32 w_nr;
92 u32 w_alloc_ptr;
93 u32 w_alloc_ctr;
94 u32 w_free_ptr;
95 atomic_t w_free_ctr;
98 /* Rings are posted with all the allocations they'll need to queue the
99 * incoming message to the receiving socket so this can't fail.
100 * All fragments start with a header, so we can make sure we're not receiving
101 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
103 struct rds_ib_ack_state {
104 u64 ack_next;
105 u64 ack_recv;
106 unsigned int ack_required:1;
107 unsigned int ack_next_valid:1;
108 unsigned int ack_recv_valid:1;
112 struct rds_ib_device;
114 struct rds_ib_connection {
116 struct list_head ib_node;
117 struct rds_ib_device *rds_ibdev;
118 struct rds_connection *conn;
120 /* alphabet soup, IBTA style */
121 struct rdma_cm_id *i_cm_id;
122 struct ib_pd *i_pd;
123 struct ib_cq *i_send_cq;
124 struct ib_cq *i_recv_cq;
125 struct ib_wc i_send_wc[RDS_IB_WC_MAX];
126 struct ib_wc i_recv_wc[RDS_IB_WC_MAX];
128 /* To control the number of wrs from fastreg */
129 atomic_t i_fastreg_wrs;
130 atomic_t i_fastunreg_wrs;
132 /* interrupt handling */
133 struct tasklet_struct i_send_tasklet;
134 struct tasklet_struct i_recv_tasklet;
136 /* tx */
137 struct rds_ib_work_ring i_send_ring;
138 struct rm_data_op *i_data_op;
139 struct rds_header *i_send_hdrs;
140 dma_addr_t i_send_hdrs_dma;
141 struct rds_ib_send_work *i_sends;
142 atomic_t i_signaled_sends;
144 /* rx */
145 struct mutex i_recv_mutex;
146 struct rds_ib_work_ring i_recv_ring;
147 struct rds_ib_incoming *i_ibinc;
148 u32 i_recv_data_rem;
149 struct rds_header *i_recv_hdrs;
150 dma_addr_t i_recv_hdrs_dma;
151 struct rds_ib_recv_work *i_recvs;
152 u64 i_ack_recv; /* last ACK received */
153 struct rds_ib_refill_cache i_cache_incs;
154 struct rds_ib_refill_cache i_cache_frags;
155 atomic_t i_cache_allocs;
157 /* sending acks */
158 unsigned long i_ack_flags;
159 #ifdef KERNEL_HAS_ATOMIC64
160 atomic64_t i_ack_next; /* next ACK to send */
161 #else
162 spinlock_t i_ack_lock; /* protect i_ack_next */
163 u64 i_ack_next; /* next ACK to send */
164 #endif
165 struct rds_header *i_ack;
166 struct ib_send_wr i_ack_wr;
167 struct ib_sge i_ack_sge;
168 dma_addr_t i_ack_dma;
169 unsigned long i_ack_queued;
171 /* Flow control related information
173 * Our algorithm uses a pair variables that we need to access
174 * atomically - one for the send credits, and one posted
175 * recv credits we need to transfer to remote.
176 * Rather than protect them using a slow spinlock, we put both into
177 * a single atomic_t and update it using cmpxchg
179 atomic_t i_credits;
181 /* Protocol version specific information */
182 unsigned int i_flowctl:1; /* enable/disable flow ctl */
184 /* Batched completions */
185 unsigned int i_unsignaled_wrs;
187 /* Endpoint role in connection */
188 bool i_active_side;
189 atomic_t i_cq_quiesce;
191 /* Send/Recv vectors */
192 int i_scq_vector;
193 int i_rcq_vector;
196 /* This assumes that atomic_t is at least 32 bits */
197 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
198 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
199 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
200 #define IB_SET_POST_CREDITS(v) ((v) << 16)
202 struct rds_ib_ipaddr {
203 struct list_head list;
204 __be32 ipaddr;
205 struct rcu_head rcu;
208 enum {
209 RDS_IB_MR_8K_POOL,
210 RDS_IB_MR_1M_POOL,
213 struct rds_ib_device {
214 struct list_head list;
215 struct list_head ipaddr_list;
216 struct list_head conn_list;
217 struct ib_device *dev;
218 struct ib_pd *pd;
219 bool has_fmr;
220 bool has_fr;
221 bool use_fastreg;
223 unsigned int max_mrs;
224 struct rds_ib_mr_pool *mr_1m_pool;
225 struct rds_ib_mr_pool *mr_8k_pool;
226 unsigned int fmr_max_remaps;
227 unsigned int max_8k_mrs;
228 unsigned int max_1m_mrs;
229 int max_sge;
230 unsigned int max_wrs;
231 unsigned int max_initiator_depth;
232 unsigned int max_responder_resources;
233 spinlock_t spinlock; /* protect the above */
234 refcount_t refcount;
235 struct work_struct free_work;
236 int *vector_load;
239 #define ibdev_to_node(ibdev) dev_to_node((ibdev)->dev.parent)
240 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
242 /* bits for i_ack_flags */
243 #define IB_ACK_IN_FLIGHT 0
244 #define IB_ACK_REQUESTED 1
246 /* Magic WR_ID for ACKs */
247 #define RDS_IB_ACK_WR_ID (~(u64) 0)
249 struct rds_ib_statistics {
250 uint64_t s_ib_connect_raced;
251 uint64_t s_ib_listen_closed_stale;
252 uint64_t s_ib_evt_handler_call;
253 uint64_t s_ib_tasklet_call;
254 uint64_t s_ib_tx_cq_event;
255 uint64_t s_ib_tx_ring_full;
256 uint64_t s_ib_tx_throttle;
257 uint64_t s_ib_tx_sg_mapping_failure;
258 uint64_t s_ib_tx_stalled;
259 uint64_t s_ib_tx_credit_updates;
260 uint64_t s_ib_rx_cq_event;
261 uint64_t s_ib_rx_ring_empty;
262 uint64_t s_ib_rx_refill_from_cq;
263 uint64_t s_ib_rx_refill_from_thread;
264 uint64_t s_ib_rx_alloc_limit;
265 uint64_t s_ib_rx_total_frags;
266 uint64_t s_ib_rx_total_incs;
267 uint64_t s_ib_rx_credit_updates;
268 uint64_t s_ib_ack_sent;
269 uint64_t s_ib_ack_send_failure;
270 uint64_t s_ib_ack_send_delayed;
271 uint64_t s_ib_ack_send_piggybacked;
272 uint64_t s_ib_ack_received;
273 uint64_t s_ib_rdma_mr_8k_alloc;
274 uint64_t s_ib_rdma_mr_8k_free;
275 uint64_t s_ib_rdma_mr_8k_used;
276 uint64_t s_ib_rdma_mr_8k_pool_flush;
277 uint64_t s_ib_rdma_mr_8k_pool_wait;
278 uint64_t s_ib_rdma_mr_8k_pool_depleted;
279 uint64_t s_ib_rdma_mr_1m_alloc;
280 uint64_t s_ib_rdma_mr_1m_free;
281 uint64_t s_ib_rdma_mr_1m_used;
282 uint64_t s_ib_rdma_mr_1m_pool_flush;
283 uint64_t s_ib_rdma_mr_1m_pool_wait;
284 uint64_t s_ib_rdma_mr_1m_pool_depleted;
285 uint64_t s_ib_rdma_mr_8k_reused;
286 uint64_t s_ib_rdma_mr_1m_reused;
287 uint64_t s_ib_atomic_cswp;
288 uint64_t s_ib_atomic_fadd;
289 uint64_t s_ib_recv_added_to_cache;
290 uint64_t s_ib_recv_removed_from_cache;
293 extern struct workqueue_struct *rds_ib_wq;
296 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
297 * doesn't define it.
299 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
300 struct scatterlist *sglist,
301 unsigned int sg_dma_len,
302 int direction)
304 struct scatterlist *sg;
305 unsigned int i;
307 for_each_sg(sglist, sg, sg_dma_len, i) {
308 ib_dma_sync_single_for_cpu(dev,
309 ib_sg_dma_address(dev, sg),
310 ib_sg_dma_len(dev, sg),
311 direction);
314 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
316 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
317 struct scatterlist *sglist,
318 unsigned int sg_dma_len,
319 int direction)
321 struct scatterlist *sg;
322 unsigned int i;
324 for_each_sg(sglist, sg, sg_dma_len, i) {
325 ib_dma_sync_single_for_device(dev,
326 ib_sg_dma_address(dev, sg),
327 ib_sg_dma_len(dev, sg),
328 direction);
331 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
334 /* ib.c */
335 extern struct rds_transport rds_ib_transport;
336 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
337 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
338 extern struct ib_client rds_ib_client;
340 extern unsigned int rds_ib_retry_count;
342 extern spinlock_t ib_nodev_conns_lock;
343 extern struct list_head ib_nodev_conns;
345 /* ib_cm.c */
346 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
347 void rds_ib_conn_free(void *arg);
348 int rds_ib_conn_path_connect(struct rds_conn_path *cp);
349 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp);
350 void rds_ib_state_change(struct sock *sk);
351 int rds_ib_listen_init(void);
352 void rds_ib_listen_stop(void);
353 __printf(2, 3)
354 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
355 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
356 struct rdma_cm_event *event);
357 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
358 void rds_ib_cm_connect_complete(struct rds_connection *conn,
359 struct rdma_cm_event *event);
362 #define rds_ib_conn_error(conn, fmt...) \
363 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
365 /* ib_rdma.c */
366 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
367 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
368 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
369 void rds_ib_destroy_nodev_conns(void);
370 void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
372 /* ib_recv.c */
373 int rds_ib_recv_init(void);
374 void rds_ib_recv_exit(void);
375 int rds_ib_recv_path(struct rds_conn_path *conn);
376 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp);
377 void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
378 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp);
379 void rds_ib_inc_free(struct rds_incoming *inc);
380 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
381 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc,
382 struct rds_ib_ack_state *state);
383 void rds_ib_recv_tasklet_fn(unsigned long data);
384 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
385 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
386 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
387 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
388 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
389 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
390 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required);
392 /* ib_ring.c */
393 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
394 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
395 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
396 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
397 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
398 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
399 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
400 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
401 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
402 extern wait_queue_head_t rds_ib_ring_empty_wait;
404 /* ib_send.c */
405 void rds_ib_xmit_path_complete(struct rds_conn_path *cp);
406 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
407 unsigned int hdr_off, unsigned int sg, unsigned int off);
408 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
409 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
410 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
411 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
412 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
413 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
414 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
415 u32 *adv_credits, int need_posted, int max_posted);
416 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
418 /* ib_stats.c */
419 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
420 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
421 #define rds_ib_stats_add(member, count) \
422 rds_stats_add_which(rds_ib_stats, member, count)
423 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
424 unsigned int avail);
426 /* ib_sysctl.c */
427 int rds_ib_sysctl_init(void);
428 void rds_ib_sysctl_exit(void);
429 extern unsigned long rds_ib_sysctl_max_send_wr;
430 extern unsigned long rds_ib_sysctl_max_recv_wr;
431 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
432 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
433 extern unsigned long rds_ib_sysctl_max_recv_allocation;
434 extern unsigned int rds_ib_sysctl_flow_control;
436 #endif