net/rds/iw.h

   1 #ifndef _RDS_IW_H
   2 #define _RDS_IW_H
   3
   4 #include <rdma/ib_verbs.h>
   5 #include <rdma/rdma_cm.h>
   6 #include "rds.h"
   7 #include "rdma_transport.h"
   8
   9 #define RDS_FASTREG_SIZE                20
  10 #define RDS_FASTREG_POOL_SIZE           2048
  11
  12 #define RDS_IW_MAX_SGE                  8
  13 #define RDS_IW_RECV_SGE                 2
  14
  15 #define RDS_IW_DEFAULT_RECV_WR          1024
  16 #define RDS_IW_DEFAULT_SEND_WR          256
  17
  18 #define RDS_IW_SUPPORTED_PROTOCOLS      0x00000003      /* minor versions supported */
  19
  20 extern struct list_head rds_iw_devices;
  21
  22 /*
  23  * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
  24  * try and minimize the amount of memory tied up both the device and
  25  * socket receive queues.
  26  */
  27 /* page offset of the final full frag that fits in the page */
  28 #define RDS_PAGE_LAST_OFF (((PAGE_SIZE  / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE)
  29 struct rds_page_frag {
  30         struct list_head        f_item;
  31         struct page             *f_page;
  32         unsigned long           f_offset;
  33         dma_addr_t              f_mapped;
  34 };
  35
  36 struct rds_iw_incoming {
  37         struct list_head        ii_frags;
  38         struct rds_incoming     ii_inc;
  39 };
  40
  41 struct rds_iw_connect_private {
  42         /* Add new fields at the end, and don't permute existing fields. */
  43         __be32                  dp_saddr;
  44         __be32                  dp_daddr;
  45         u8                      dp_protocol_major;
  46         u8                      dp_protocol_minor;
  47         __be16                  dp_protocol_minor_mask; /* bitmask */
  48         __be32                  dp_reserved1;
  49         __be64                  dp_ack_seq;
  50         __be32                  dp_credit;              /* non-zero enables flow ctl */
  51 };
  52
  53 struct rds_iw_scatterlist {
  54         struct scatterlist      *list;
  55         unsigned int            len;
  56         int                     dma_len;
  57         unsigned int            dma_npages;
  58         unsigned int            bytes;
  59 };
  60
  61 struct rds_iw_mapping {
  62         spinlock_t              m_lock; /* protect the mapping struct */
  63         struct list_head        m_list;
  64         struct rds_iw_mr        *m_mr;
  65         uint32_t                m_rkey;
  66         struct rds_iw_scatterlist m_sg;
  67 };
  68
  69 struct rds_iw_send_work {
  70         struct rds_message      *s_rm;
  71
  72         /* We should really put these into a union: */
  73         struct rds_rdma_op      *s_op;
  74         struct rds_iw_mapping   *s_mapping;
  75         struct ib_mr            *s_mr;
  76         struct ib_fast_reg_page_list *s_page_list;
  77         unsigned char           s_remap_count;
  78
  79         struct ib_send_wr       s_wr;
  80         struct ib_sge           s_sge[RDS_IW_MAX_SGE];
  81         unsigned long           s_queued;
  82 };
  83
  84 struct rds_iw_recv_work {
  85         struct rds_iw_incoming  *r_iwinc;
  86         struct rds_page_frag    *r_frag;
  87         struct ib_recv_wr       r_wr;
  88         struct ib_sge           r_sge[2];
  89 };
  90
  91 struct rds_iw_work_ring {
  92         u32             w_nr;
  93         u32             w_alloc_ptr;
  94         u32             w_alloc_ctr;
  95         u32             w_free_ptr;
  96         atomic_t        w_free_ctr;
  97 };
  98
  99 struct rds_iw_device;
 100
 101 struct rds_iw_connection {
 102
 103         struct list_head        iw_node;
 104         struct rds_iw_device    *rds_iwdev;
 105         struct rds_connection   *conn;
 106
 107         /* alphabet soup, IBTA style */
 108         struct rdma_cm_id       *i_cm_id;
 109         struct ib_pd            *i_pd;
 110         struct ib_mr            *i_mr;
 111         struct ib_cq            *i_send_cq;
 112         struct ib_cq            *i_recv_cq;
 113
 114         /* tx */
 115         struct rds_iw_work_ring i_send_ring;
 116         struct rds_message      *i_rm;
 117         struct rds_header       *i_send_hdrs;
 118         u64                     i_send_hdrs_dma;
 119         struct rds_iw_send_work *i_sends;
 120
 121         /* rx */
 122         struct mutex            i_recv_mutex;
 123         struct rds_iw_work_ring i_recv_ring;
 124         struct rds_iw_incoming  *i_iwinc;
 125         u32                     i_recv_data_rem;
 126         struct rds_header       *i_recv_hdrs;
 127         u64                     i_recv_hdrs_dma;
 128         struct rds_iw_recv_work *i_recvs;
 129         struct rds_page_frag    i_frag;
 130         u64                     i_ack_recv;     /* last ACK received */
 131
 132         /* sending acks */
 133         unsigned long           i_ack_flags;
 134         u64                     i_ack_next;     /* next ACK to send */
 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         unsigned int            i_dma_local_lkey:1;
 154         unsigned int            i_fastreg_posted:1; /* fastreg posted on this connection */
 155         /* Batched completions */
 156         unsigned int            i_unsignaled_wrs;
 157         long                    i_unsignaled_bytes;
 158 };
 159
 160 /* This assumes that atomic_t is at least 32 bits */
 161 #define IB_GET_SEND_CREDITS(v)  ((v) & 0xffff)
 162 #define IB_GET_POST_CREDITS(v)  ((v) >> 16)
 163 #define IB_SET_SEND_CREDITS(v)  ((v) & 0xffff)
 164 #define IB_SET_POST_CREDITS(v)  ((v) << 16)
 165
 166 struct rds_iw_cm_id {
 167         struct list_head        list;
 168         struct rdma_cm_id       *cm_id;
 169 };
 170
 171 struct rds_iw_device {
 172         struct list_head        list;
 173         struct list_head        cm_id_list;
 174         struct list_head        conn_list;
 175         struct ib_device        *dev;
 176         struct ib_pd            *pd;
 177         struct ib_mr            *mr;
 178         struct rds_iw_mr_pool   *mr_pool;
 179         int                     page_shift;
 180         int                     max_sge;
 181         unsigned int            max_wrs;
 182         unsigned int            dma_local_lkey:1;
 183         spinlock_t              spinlock;       /* protect the above */
 184 };
 185
 186 /* bits for i_ack_flags */
 187 #define IB_ACK_IN_FLIGHT        0
 188 #define IB_ACK_REQUESTED        1
 189
 190 /* Magic WR_ID for ACKs */
 191 #define RDS_IW_ACK_WR_ID        ((u64)0xffffffffffffffffULL)
 192 #define RDS_IW_FAST_REG_WR_ID   ((u64)0xefefefefefefefefULL)
 193 #define RDS_IW_LOCAL_INV_WR_ID  ((u64)0xdfdfdfdfdfdfdfdfULL)
 194
 195 struct rds_iw_statistics {
 196         uint64_t        s_iw_connect_raced;
 197         uint64_t        s_iw_listen_closed_stale;
 198         uint64_t        s_iw_tx_cq_call;
 199         uint64_t        s_iw_tx_cq_event;
 200         uint64_t        s_iw_tx_ring_full;
 201         uint64_t        s_iw_tx_throttle;
 202         uint64_t        s_iw_tx_sg_mapping_failure;
 203         uint64_t        s_iw_tx_stalled;
 204         uint64_t        s_iw_tx_credit_updates;
 205         uint64_t        s_iw_rx_cq_call;
 206         uint64_t        s_iw_rx_cq_event;
 207         uint64_t        s_iw_rx_ring_empty;
 208         uint64_t        s_iw_rx_refill_from_cq;
 209         uint64_t        s_iw_rx_refill_from_thread;
 210         uint64_t        s_iw_rx_alloc_limit;
 211         uint64_t        s_iw_rx_credit_updates;
 212         uint64_t        s_iw_ack_sent;
 213         uint64_t        s_iw_ack_send_failure;
 214         uint64_t        s_iw_ack_send_delayed;
 215         uint64_t        s_iw_ack_send_piggybacked;
 216         uint64_t        s_iw_ack_received;
 217         uint64_t        s_iw_rdma_mr_alloc;
 218         uint64_t        s_iw_rdma_mr_free;
 219         uint64_t        s_iw_rdma_mr_used;
 220         uint64_t        s_iw_rdma_mr_pool_flush;
 221         uint64_t        s_iw_rdma_mr_pool_wait;
 222         uint64_t        s_iw_rdma_mr_pool_depleted;
 223 };
 224
 225 extern struct workqueue_struct *rds_iw_wq;
 226
 227 /*
 228  * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
 229  * doesn't define it.
 230  */
 231 static inline void rds_iw_dma_sync_sg_for_cpu(struct ib_device *dev,
 232                 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
 233 {
 234         unsigned int i;
 235
 236         for (i = 0; i < sg_dma_len; ++i) {
 237                 ib_dma_sync_single_for_cpu(dev,
 238                                 ib_sg_dma_address(dev, &sg[i]),
 239                                 ib_sg_dma_len(dev, &sg[i]),
 240                                 direction);
 241         }
 242 }
 243 #define ib_dma_sync_sg_for_cpu  rds_iw_dma_sync_sg_for_cpu
 244
 245 static inline void rds_iw_dma_sync_sg_for_device(struct ib_device *dev,
 246                 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
 247 {
 248         unsigned int i;
 249
 250         for (i = 0; i < sg_dma_len; ++i) {
 251                 ib_dma_sync_single_for_device(dev,
 252                                 ib_sg_dma_address(dev, &sg[i]),
 253                                 ib_sg_dma_len(dev, &sg[i]),
 254                                 direction);
 255         }
 256 }
 257 #define ib_dma_sync_sg_for_device       rds_iw_dma_sync_sg_for_device
 258
 259 static inline u32 rds_iw_local_dma_lkey(struct rds_iw_connection *ic)
 260 {
 261         return ic->i_dma_local_lkey ? ic->i_cm_id->device->local_dma_lkey : ic->i_mr->lkey;
 262 }
 263
 264 /* ib.c */
 265 extern struct rds_transport rds_iw_transport;
 266 extern void rds_iw_add_one(struct ib_device *device);
 267 extern void rds_iw_remove_one(struct ib_device *device);
 268 extern struct ib_client rds_iw_client;
 269
 270 extern unsigned int fastreg_pool_size;
 271 extern unsigned int fastreg_message_size;
 272
 273 extern spinlock_t iw_nodev_conns_lock;
 274 extern struct list_head iw_nodev_conns;
 275
 276 /* ib_cm.c */
 277 int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp);
 278 void rds_iw_conn_free(void *arg);
 279 int rds_iw_conn_connect(struct rds_connection *conn);
 280 void rds_iw_conn_shutdown(struct rds_connection *conn);
 281 void rds_iw_state_change(struct sock *sk);
 282 int __init rds_iw_listen_init(void);
 283 void rds_iw_listen_stop(void);
 284 void __rds_iw_conn_error(struct rds_connection *conn, const char *, ...);
 285 int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
 286                              struct rdma_cm_event *event);
 287 int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id);
 288 void rds_iw_cm_connect_complete(struct rds_connection *conn,
 289                                 struct rdma_cm_event *event);
 290
 291
 292 #define rds_iw_conn_error(conn, fmt...) \
 293         __rds_iw_conn_error(conn, KERN_WARNING "RDS/IW: " fmt)
 294
 295 /* ib_rdma.c */
 296 int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
 297 int rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
 298 void rds_iw_remove_nodev_conns(void);
 299 void rds_iw_remove_conns(struct rds_iw_device *rds_iwdev);
 300 struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *);
 301 void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo);
 302 void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *);
 303 void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
 304                     struct rds_sock *rs, u32 *key_ret);
 305 void rds_iw_sync_mr(void *trans_private, int dir);
 306 void rds_iw_free_mr(void *trans_private, int invalidate);
 307 void rds_iw_flush_mrs(void);
 308 void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
 309
 310 /* ib_recv.c */
 311 int __init rds_iw_recv_init(void);
 312 void rds_iw_recv_exit(void);
 313 int rds_iw_recv(struct rds_connection *conn);
 314 int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
 315                        gfp_t page_gfp, int prefill);
 316 void rds_iw_inc_purge(struct rds_incoming *inc);
 317 void rds_iw_inc_free(struct rds_incoming *inc);
 318 int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
 319                              size_t size);
 320 void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context);
 321 void rds_iw_recv_init_ring(struct rds_iw_connection *ic);
 322 void rds_iw_recv_clear_ring(struct rds_iw_connection *ic);
 323 void rds_iw_recv_init_ack(struct rds_iw_connection *ic);
 324 void rds_iw_attempt_ack(struct rds_iw_connection *ic);
 325 void rds_iw_ack_send_complete(struct rds_iw_connection *ic);
 326 u64 rds_iw_piggyb_ack(struct rds_iw_connection *ic);
 327
 328 /* ib_ring.c */
 329 void rds_iw_ring_init(struct rds_iw_work_ring *ring, u32 nr);
 330 void rds_iw_ring_resize(struct rds_iw_work_ring *ring, u32 nr);
 331 u32 rds_iw_ring_alloc(struct rds_iw_work_ring *ring, u32 val, u32 *pos);
 332 void rds_iw_ring_free(struct rds_iw_work_ring *ring, u32 val);
 333 void rds_iw_ring_unalloc(struct rds_iw_work_ring *ring, u32 val);
 334 int rds_iw_ring_empty(struct rds_iw_work_ring *ring);
 335 int rds_iw_ring_low(struct rds_iw_work_ring *ring);
 336 u32 rds_iw_ring_oldest(struct rds_iw_work_ring *ring);
 337 u32 rds_iw_ring_completed(struct rds_iw_work_ring *ring, u32 wr_id, u32 oldest);
 338 extern wait_queue_head_t rds_iw_ring_empty_wait;
 339
 340 /* ib_send.c */
 341 void rds_iw_xmit_complete(struct rds_connection *conn);
 342 int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 343                 unsigned int hdr_off, unsigned int sg, unsigned int off);
 344 void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context);
 345 void rds_iw_send_init_ring(struct rds_iw_connection *ic);
 346 void rds_iw_send_clear_ring(struct rds_iw_connection *ic);
 347 int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op);
 348 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits);
 349 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted);
 350 int rds_iw_send_grab_credits(struct rds_iw_connection *ic, u32 wanted,
 351                              u32 *adv_credits, int need_posted);
 352
 353 /* ib_stats.c */
 354 DECLARE_PER_CPU(struct rds_iw_statistics, rds_iw_stats);
 355 #define rds_iw_stats_inc(member) rds_stats_inc_which(rds_iw_stats, member)
 356 unsigned int rds_iw_stats_info_copy(struct rds_info_iterator *iter,
 357                                     unsigned int avail);
 358
 359 /* ib_sysctl.c */
 360 int __init rds_iw_sysctl_init(void);
 361 void rds_iw_sysctl_exit(void);
 362 extern unsigned long rds_iw_sysctl_max_send_wr;
 363 extern unsigned long rds_iw_sysctl_max_recv_wr;
 364 extern unsigned long rds_iw_sysctl_max_unsig_wrs;
 365 extern unsigned long rds_iw_sysctl_max_unsig_bytes;
 366 extern unsigned long rds_iw_sysctl_max_recv_allocation;
 367 extern unsigned int rds_iw_sysctl_flow_control;
 368 extern ctl_table rds_iw_sysctl_table[];
 369
 370 /*
 371  * Helper functions for getting/setting the header and data SGEs in
 372  * RDS packets (not RDMA)
 373  */
 374 static inline struct ib_sge *
 375 rds_iw_header_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
 376 {
 377         return &sge[0];
 378 }
 379
 380 static inline struct ib_sge *
 381 rds_iw_data_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
 382 {
 383         return &sge[1];
 384 }
 385
 386 static inline void rds_iw_set_64bit(u64 *ptr, u64 val)
 387 {
 388 #if BITS_PER_LONG == 64
 389         *ptr = val;
 390 #else
 391         set_64bit(ptr, val);
 392 #endif
 393 }
 394
 395 #endif