net/rds/ib.h

   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/pci.h>
   7 #include <linux/slab.h>
   8 #include "rds.h"
   9 #include "rdma_transport.h"
  10
  11 #define RDS_FMR_SIZE                    256
  12 #define RDS_FMR_POOL_SIZE               8192
  13
  14 #define RDS_IB_MAX_SGE                  8
  15 #define RDS_IB_RECV_SGE                 2
  16
  17 #define RDS_IB_DEFAULT_RECV_WR          1024
  18 #define RDS_IB_DEFAULT_SEND_WR          256
  19
  20 #define RDS_IB_DEFAULT_RETRY_COUNT      2
  21
  22 #define RDS_IB_SUPPORTED_PROTOCOLS      0x00000003      /* minor versions supported */
  23
  24 #define RDS_IB_RECYCLE_BATCH_COUNT      32
  25
  26 extern struct rw_semaphore rds_ib_devices_lock;
  27 extern struct list_head rds_ib_devices;
  28
  29 /*
  30  * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
  31  * try and minimize the amount of memory tied up both the device and
  32  * socket receive queues.
  33  */
  34 struct rds_page_frag {
  35         struct list_head        f_item;
  36         struct list_head        f_cache_entry;
  37         struct scatterlist      f_sg;
  38 };
  39
  40 struct rds_ib_incoming {
  41         struct list_head        ii_frags;
  42         struct list_head        ii_cache_entry;
  43         struct rds_incoming     ii_inc;
  44 };
  45
  46 struct rds_ib_cache_head {
  47         struct list_head *first;
  48         unsigned long count;
  49 };
  50
  51 struct rds_ib_refill_cache {
  52         struct rds_ib_cache_head *percpu;
  53         struct list_head         *xfer;
  54         struct list_head         *ready;
  55 };
  56
  57 struct rds_ib_connect_private {
  58         /* Add new fields at the end, and don't permute existing fields. */
  59         __be32                  dp_saddr;
  60         __be32                  dp_daddr;
  61         u8                      dp_protocol_major;
  62         u8                      dp_protocol_minor;
  63         __be16                  dp_protocol_minor_mask; /* bitmask */
  64         __be32                  dp_reserved1;
  65         __be64                  dp_ack_seq;
  66         __be32                  dp_credit;              /* non-zero enables flow ctl */
  67 };
  68
  69 struct rds_ib_send_work {
  70         void                    *s_op;
  71         struct ib_send_wr       s_wr;
  72         struct ib_sge           s_sge[RDS_IB_MAX_SGE];
  73         unsigned long           s_queued;
  74 };
  75
  76 struct rds_ib_recv_work {
  77         struct rds_ib_incoming  *r_ibinc;
  78         struct rds_page_frag    *r_frag;
  79         struct ib_recv_wr       r_wr;
  80         struct ib_sge           r_sge[2];
  81 };
  82
  83 struct rds_ib_work_ring {
  84         u32             w_nr;
  85         u32             w_alloc_ptr;
  86         u32             w_alloc_ctr;
  87         u32             w_free_ptr;
  88         atomic_t        w_free_ctr;
  89 };
  90
  91 struct rds_ib_device;
  92
  93 struct rds_ib_connection {
  94
  95         struct list_head        ib_node;
  96         struct rds_ib_device    *rds_ibdev;
  97         struct rds_connection   *conn;
  98
  99         /* alphabet soup, IBTA style */
 100         struct rdma_cm_id       *i_cm_id;
 101         struct ib_pd            *i_pd;
 102         struct ib_mr            *i_mr;
 103         struct ib_cq            *i_send_cq;
 104         struct ib_cq            *i_recv_cq;
 105
 106         /* tx */
 107         struct rds_ib_work_ring i_send_ring;
 108         struct rm_data_op       *i_data_op;
 109         struct rds_header       *i_send_hdrs;
 110         u64                     i_send_hdrs_dma;
 111         struct rds_ib_send_work *i_sends;
 112         atomic_t                i_signaled_sends;
 113
 114         /* rx */
 115         struct tasklet_struct   i_recv_tasklet;
 116         struct mutex            i_recv_mutex;
 117         struct rds_ib_work_ring i_recv_ring;
 118         struct rds_ib_incoming  *i_ibinc;
 119         u32                     i_recv_data_rem;
 120         struct rds_header       *i_recv_hdrs;
 121         u64                     i_recv_hdrs_dma;
 122         struct rds_ib_recv_work *i_recvs;
 123         u64                     i_ack_recv;     /* last ACK received */
 124         struct rds_ib_refill_cache i_cache_incs;
 125         struct rds_ib_refill_cache i_cache_frags;
 126
 127         /* sending acks */
 128         unsigned long           i_ack_flags;
 129 #ifdef KERNEL_HAS_ATOMIC64
 130         atomic64_t              i_ack_next;     /* next ACK to send */
 131 #else
 132         spinlock_t              i_ack_lock;     /* protect i_ack_next */
 133         u64                     i_ack_next;     /* next ACK to send */
 134 #endif
 135         struct rds_header       *i_ack;
 136         struct ib_send_wr       i_ack_wr;
 137         struct ib_sge           i_ack_sge;
 138         u64                     i_ack_dma;
 139         unsigned long           i_ack_queued;
 140
 141         /* Flow control related information
 142          *
 143          * Our algorithm uses a pair variables that we need to access
 144          * atomically - one for the send credits, and one posted
 145          * recv credits we need to transfer to remote.
 146          * Rather than protect them using a slow spinlock, we put both into
 147          * a single atomic_t and update it using cmpxchg
 148          */
 149         atomic_t                i_credits;
 150
 151         /* Protocol version specific information */
 152         unsigned int            i_flowctl:1;    /* enable/disable flow ctl */
 153
 154         /* Batched completions */
 155         unsigned int            i_unsignaled_wrs;
 156 };
 157
 158 /* This assumes that atomic_t is at least 32 bits */
 159 #define IB_GET_SEND_CREDITS(v)  ((v) & 0xffff)
 160 #define IB_GET_POST_CREDITS(v)  ((v) >> 16)
 161 #define IB_SET_SEND_CREDITS(v)  ((v) & 0xffff)
 162 #define IB_SET_POST_CREDITS(v)  ((v) << 16)
 163
 164 struct rds_ib_ipaddr {
 165         struct list_head        list;
 166         __be32                  ipaddr;
 167 };
 168
 169 struct rds_ib_device {
 170         struct list_head        list;
 171         struct list_head        ipaddr_list;
 172         struct list_head        conn_list;
 173         struct ib_device        *dev;
 174         struct ib_pd            *pd;
 175         struct ib_mr            *mr;
 176         struct rds_ib_mr_pool   *mr_pool;
 177         unsigned int            fmr_max_remaps;
 178         unsigned int            max_fmrs;
 179         int                     max_sge;
 180         unsigned int            max_wrs;
 181         unsigned int            max_initiator_depth;
 182         unsigned int            max_responder_resources;
 183         spinlock_t              spinlock;       /* protect the above */
 184         atomic_t                refcount;
 185         struct work_struct      free_work;
 186 };
 187
 188 #define pcidev_to_node(pcidev) pcibus_to_node(pcidev->bus)
 189 #define ibdev_to_node(ibdev) pcidev_to_node(to_pci_dev(ibdev->dma_device))
 190 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
 191
 192 /* bits for i_ack_flags */
 193 #define IB_ACK_IN_FLIGHT        0
 194 #define IB_ACK_REQUESTED        1
 195
 196 /* Magic WR_ID for ACKs */
 197 #define RDS_IB_ACK_WR_ID        (~(u64) 0)
 198
 199 struct rds_ib_statistics {
 200         uint64_t        s_ib_connect_raced;
 201         uint64_t        s_ib_listen_closed_stale;
 202         uint64_t        s_ib_tx_cq_call;
 203         uint64_t        s_ib_tx_cq_event;
 204         uint64_t        s_ib_tx_ring_full;
 205         uint64_t        s_ib_tx_throttle;
 206         uint64_t        s_ib_tx_sg_mapping_failure;
 207         uint64_t        s_ib_tx_stalled;
 208         uint64_t        s_ib_tx_credit_updates;
 209         uint64_t        s_ib_rx_cq_call;
 210         uint64_t        s_ib_rx_cq_event;
 211         uint64_t        s_ib_rx_ring_empty;
 212         uint64_t        s_ib_rx_refill_from_cq;
 213         uint64_t        s_ib_rx_refill_from_thread;
 214         uint64_t        s_ib_rx_alloc_limit;
 215         uint64_t        s_ib_rx_credit_updates;
 216         uint64_t        s_ib_ack_sent;
 217         uint64_t        s_ib_ack_send_failure;
 218         uint64_t        s_ib_ack_send_delayed;
 219         uint64_t        s_ib_ack_send_piggybacked;
 220         uint64_t        s_ib_ack_received;
 221         uint64_t        s_ib_rdma_mr_alloc;
 222         uint64_t        s_ib_rdma_mr_free;
 223         uint64_t        s_ib_rdma_mr_used;
 224         uint64_t        s_ib_rdma_mr_pool_flush;
 225         uint64_t        s_ib_rdma_mr_pool_wait;
 226         uint64_t        s_ib_rdma_mr_pool_depleted;
 227         uint64_t        s_ib_atomic_cswp;
 228         uint64_t        s_ib_atomic_fadd;
 229 };
 230
 231 extern struct workqueue_struct *rds_ib_wq;
 232
 233 /*
 234  * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
 235  * doesn't define it.
 236  */
 237 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
 238                 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
 239 {
 240         unsigned int i;
 241
 242         for (i = 0; i < sg_dma_len; ++i) {
 243                 ib_dma_sync_single_for_cpu(dev,
 244                                 ib_sg_dma_address(dev, &sg[i]),
 245                                 ib_sg_dma_len(dev, &sg[i]),
 246                                 direction);
 247         }
 248 }
 249 #define ib_dma_sync_sg_for_cpu  rds_ib_dma_sync_sg_for_cpu
 250
 251 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
 252                 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
 253 {
 254         unsigned int i;
 255
 256         for (i = 0; i < sg_dma_len; ++i) {
 257                 ib_dma_sync_single_for_device(dev,
 258                                 ib_sg_dma_address(dev, &sg[i]),
 259                                 ib_sg_dma_len(dev, &sg[i]),
 260                                 direction);
 261         }
 262 }
 263 #define ib_dma_sync_sg_for_device       rds_ib_dma_sync_sg_for_device
 264
 265
 266 /* ib.c */
 267 extern struct rds_transport rds_ib_transport;
 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;
 271
 272 extern unsigned int fmr_message_size;
 273 extern unsigned int rds_ib_retry_count;
 274
 275 extern spinlock_t ib_nodev_conns_lock;
 276 extern struct list_head ib_nodev_conns;
 277
 278 /* ib_cm.c */
 279 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
 280 void rds_ib_conn_free(void *arg);
 281 int rds_ib_conn_connect(struct rds_connection *conn);
 282 void rds_ib_conn_shutdown(struct rds_connection *conn);
 283 void rds_ib_state_change(struct sock *sk);
 284 int rds_ib_listen_init(void);
 285 void rds_ib_listen_stop(void);
 286 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
 287 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 288                              struct rdma_cm_event *event);
 289 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
 290 void rds_ib_cm_connect_complete(struct rds_connection *conn,
 291                                 struct rdma_cm_event *event);
 292
 293
 294 #define rds_ib_conn_error(conn, fmt...) \
 295         __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
 296
 297 /* ib_rdma.c */
 298 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
 299 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
 300 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
 301 void rds_ib_destroy_nodev_conns(void);
 302 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
 303 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
 304 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
 305 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 306                     struct rds_sock *rs, u32 *key_ret);
 307 void rds_ib_sync_mr(void *trans_private, int dir);
 308 void rds_ib_free_mr(void *trans_private, int invalidate);
 309 void rds_ib_flush_mrs(void);
 310
 311 /* ib_recv.c */
 312 int rds_ib_recv_init(void);
 313 void rds_ib_recv_exit(void);
 314 int rds_ib_recv(struct rds_connection *conn);
 315 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic);
 316 void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
 317 void rds_ib_recv_refill(struct rds_connection *conn, int prefill);
 318 void rds_ib_inc_free(struct rds_incoming *inc);
 319 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
 320                              size_t size);
 321 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
 322 void rds_ib_recv_tasklet_fn(unsigned long data);
 323 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
 324 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
 325 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
 326 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
 327 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
 328 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
 329
 330 /* ib_ring.c */
 331 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
 332 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
 333 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
 334 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
 335 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
 336 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
 337 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
 338 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
 339 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
 340 extern wait_queue_head_t rds_ib_ring_empty_wait;
 341
 342 /* ib_send.c */
 343 char *rds_ib_wc_status_str(enum ib_wc_status status);
 344 void rds_ib_xmit_complete(struct rds_connection *conn);
 345 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 346                 unsigned int hdr_off, unsigned int sg, unsigned int off);
 347 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
 348 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
 349 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
 350 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
 351 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
 352 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
 353 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
 354                              u32 *adv_credits, int need_posted, int max_posted);
 355 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
 356
 357 /* ib_stats.c */
 358 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
 359 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
 360 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
 361                                     unsigned int avail);
 362
 363 /* ib_sysctl.c */
 364 int rds_ib_sysctl_init(void);
 365 void rds_ib_sysctl_exit(void);
 366 extern unsigned long rds_ib_sysctl_max_send_wr;
 367 extern unsigned long rds_ib_sysctl_max_recv_wr;
 368 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
 369 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
 370 extern unsigned long rds_ib_sysctl_max_recv_allocation;
 371 extern unsigned int rds_ib_sysctl_flow_control;
 372
 373 #endif