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               4096
  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 extern struct list_head rds_ib_devices;
  25
  26 /*
  27  * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
  28  * try and minimize the amount of memory tied up both the device and
  29  * socket receive queues.
  30  */
  31 struct rds_page_frag {
  32         struct list_head        f_item;
  33         struct scatterlist      f_sg;
  34 };
  35
  36 struct rds_ib_incoming {
  37         struct list_head        ii_frags;
  38         struct rds_incoming     ii_inc;
  39 };
  40
  41 struct rds_ib_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_ib_send_work {
  54         void                    *s_op;
  55         struct ib_send_wr       s_wr;
  56         struct ib_sge           s_sge[RDS_IB_MAX_SGE];
  57         unsigned long           s_queued;
  58 };
  59
  60 struct rds_ib_recv_work {
  61         struct rds_ib_incoming  *r_ibinc;
  62         struct rds_page_frag    *r_frag;
  63         struct ib_recv_wr       r_wr;
  64         struct ib_sge           r_sge[2];
  65 };
  66
  67 struct rds_ib_work_ring {
  68         u32             w_nr;
  69         u32             w_alloc_ptr;
  70         u32             w_alloc_ctr;
  71         u32             w_free_ptr;
  72         atomic_t        w_free_ctr;
  73 };
  74
  75 struct rds_ib_device;
  76
  77 struct rds_ib_connection {
  78
  79         struct list_head        ib_node;
  80         struct rds_ib_device    *rds_ibdev;
  81         struct rds_connection   *conn;
  82
  83         /* alphabet soup, IBTA style */
  84         struct rdma_cm_id       *i_cm_id;
  85         struct ib_pd            *i_pd;
  86         struct ib_mr            *i_mr;
  87         struct ib_cq            *i_send_cq;
  88         struct ib_cq            *i_recv_cq;
  89
  90         /* tx */
  91         struct rds_ib_work_ring i_send_ring;
  92         struct rm_data_op       *i_data_op;
  93         struct rds_header       *i_send_hdrs;
  94         u64                     i_send_hdrs_dma;
  95         struct rds_ib_send_work *i_sends;
  96
  97         /* rx */
  98         struct tasklet_struct   i_recv_tasklet;
  99         struct mutex            i_recv_mutex;
 100         struct rds_ib_work_ring i_recv_ring;
 101         struct rds_ib_incoming  *i_ibinc;
 102         u32                     i_recv_data_rem;
 103         struct rds_header       *i_recv_hdrs;
 104         u64                     i_recv_hdrs_dma;
 105         struct rds_ib_recv_work *i_recvs;
 106         u64                     i_ack_recv;     /* last ACK received */
 107
 108         /* sending acks */
 109         unsigned long           i_ack_flags;
 110 #ifdef KERNEL_HAS_ATOMIC64
 111         atomic64_t              i_ack_next;     /* next ACK to send */
 112 #else
 113         spinlock_t              i_ack_lock;     /* protect i_ack_next */
 114         u64                     i_ack_next;     /* next ACK to send */
 115 #endif
 116         struct rds_header       *i_ack;
 117         struct ib_send_wr       i_ack_wr;
 118         struct ib_sge           i_ack_sge;
 119         u64                     i_ack_dma;
 120         unsigned long           i_ack_queued;
 121
 122         /* Flow control related information
 123          *
 124          * Our algorithm uses a pair variables that we need to access
 125          * atomically - one for the send credits, and one posted
 126          * recv credits we need to transfer to remote.
 127          * Rather than protect them using a slow spinlock, we put both into
 128          * a single atomic_t and update it using cmpxchg
 129          */
 130         atomic_t                i_credits;
 131
 132         /* Protocol version specific information */
 133         unsigned int            i_flowctl:1;    /* enable/disable flow ctl */
 134
 135         /* Batched completions */
 136         unsigned int            i_unsignaled_wrs;
 137 };
 138
 139 /* This assumes that atomic_t is at least 32 bits */
 140 #define IB_GET_SEND_CREDITS(v)  ((v) & 0xffff)
 141 #define IB_GET_POST_CREDITS(v)  ((v) >> 16)
 142 #define IB_SET_SEND_CREDITS(v)  ((v) & 0xffff)
 143 #define IB_SET_POST_CREDITS(v)  ((v) << 16)
 144
 145 struct rds_ib_ipaddr {
 146         struct list_head        list;
 147         __be32                  ipaddr;
 148 };
 149
 150 struct rds_ib_device {
 151         struct list_head        list;
 152         struct list_head        ipaddr_list;
 153         struct list_head        conn_list;
 154         struct ib_device        *dev;
 155         struct ib_pd            *pd;
 156         struct ib_mr            *mr;
 157         struct rds_ib_mr_pool   *mr_pool;
 158         unsigned int            fmr_max_remaps;
 159         unsigned int            max_fmrs;
 160         int                     max_sge;
 161         unsigned int            max_wrs;
 162         unsigned int            max_initiator_depth;
 163         unsigned int            max_responder_resources;
 164         spinlock_t              spinlock;       /* protect the above */
 165         atomic_t                refcount;
 166         struct work_struct      free_work;
 167 };
 168
 169 #define pcidev_to_node(pcidev) pcibus_to_node(pcidev->bus)
 170 #define ibdev_to_node(ibdev) pcidev_to_node(to_pci_dev(ibdev->dma_device))
 171 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
 172
 173 /* bits for i_ack_flags */
 174 #define IB_ACK_IN_FLIGHT        0
 175 #define IB_ACK_REQUESTED        1
 176
 177 /* Magic WR_ID for ACKs */
 178 #define RDS_IB_ACK_WR_ID        (~(u64) 0)
 179
 180 struct rds_ib_statistics {
 181         uint64_t        s_ib_connect_raced;
 182         uint64_t        s_ib_listen_closed_stale;
 183         uint64_t        s_ib_tx_cq_call;
 184         uint64_t        s_ib_tx_cq_event;
 185         uint64_t        s_ib_tx_ring_full;
 186         uint64_t        s_ib_tx_throttle;
 187         uint64_t        s_ib_tx_sg_mapping_failure;
 188         uint64_t        s_ib_tx_stalled;
 189         uint64_t        s_ib_tx_credit_updates;
 190         uint64_t        s_ib_rx_cq_call;
 191         uint64_t        s_ib_rx_cq_event;
 192         uint64_t        s_ib_rx_ring_empty;
 193         uint64_t        s_ib_rx_refill_from_cq;
 194         uint64_t        s_ib_rx_refill_from_thread;
 195         uint64_t        s_ib_rx_alloc_limit;
 196         uint64_t        s_ib_rx_credit_updates;
 197         uint64_t        s_ib_ack_sent;
 198         uint64_t        s_ib_ack_send_failure;
 199         uint64_t        s_ib_ack_send_delayed;
 200         uint64_t        s_ib_ack_send_piggybacked;
 201         uint64_t        s_ib_ack_received;
 202         uint64_t        s_ib_rdma_mr_alloc;
 203         uint64_t        s_ib_rdma_mr_free;
 204         uint64_t        s_ib_rdma_mr_used;
 205         uint64_t        s_ib_rdma_mr_pool_flush;
 206         uint64_t        s_ib_rdma_mr_pool_wait;
 207         uint64_t        s_ib_rdma_mr_pool_depleted;
 208         uint64_t        s_ib_atomic_cswp;
 209         uint64_t        s_ib_atomic_fadd;
 210 };
 211
 212 extern struct workqueue_struct *rds_ib_wq;
 213
 214 /*
 215  * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
 216  * doesn't define it.
 217  */
 218 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
 219                 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
 220 {
 221         unsigned int i;
 222
 223         for (i = 0; i < sg_dma_len; ++i) {
 224                 ib_dma_sync_single_for_cpu(dev,
 225                                 ib_sg_dma_address(dev, &sg[i]),
 226                                 ib_sg_dma_len(dev, &sg[i]),
 227                                 direction);
 228         }
 229 }
 230 #define ib_dma_sync_sg_for_cpu  rds_ib_dma_sync_sg_for_cpu
 231
 232 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
 233                 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
 234 {
 235         unsigned int i;
 236
 237         for (i = 0; i < sg_dma_len; ++i) {
 238                 ib_dma_sync_single_for_device(dev,
 239                                 ib_sg_dma_address(dev, &sg[i]),
 240                                 ib_sg_dma_len(dev, &sg[i]),
 241                                 direction);
 242         }
 243 }
 244 #define ib_dma_sync_sg_for_device       rds_ib_dma_sync_sg_for_device
 245
 246
 247 /* ib.c */
 248 extern struct rds_transport rds_ib_transport;
 249 extern void rds_ib_add_one(struct ib_device *device);
 250 extern void rds_ib_remove_one(struct ib_device *device);
 251 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
 252 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
 253 extern struct ib_client rds_ib_client;
 254
 255 extern unsigned int fmr_pool_size;
 256 extern unsigned int fmr_message_size;
 257 extern unsigned int rds_ib_retry_count;
 258
 259 extern spinlock_t ib_nodev_conns_lock;
 260 extern struct list_head ib_nodev_conns;
 261
 262 /* ib_cm.c */
 263 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
 264 void rds_ib_conn_free(void *arg);
 265 int rds_ib_conn_connect(struct rds_connection *conn);
 266 void rds_ib_conn_shutdown(struct rds_connection *conn);
 267 void rds_ib_state_change(struct sock *sk);
 268 int __init rds_ib_listen_init(void);
 269 void rds_ib_listen_stop(void);
 270 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
 271 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 272                              struct rdma_cm_event *event);
 273 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
 274 void rds_ib_cm_connect_complete(struct rds_connection *conn,
 275                                 struct rdma_cm_event *event);
 276
 277
 278 #define rds_ib_conn_error(conn, fmt...) \
 279         __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
 280
 281 /* ib_rdma.c */
 282 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
 283 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
 284 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
 285 void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock);
 286 static inline void rds_ib_destroy_nodev_conns(void)
 287 {
 288         __rds_ib_destroy_conns(&ib_nodev_conns, &ib_nodev_conns_lock);
 289 }
 290 static inline void rds_ib_destroy_conns(struct rds_ib_device *rds_ibdev)
 291 {
 292         __rds_ib_destroy_conns(&rds_ibdev->conn_list, &rds_ibdev->spinlock);
 293 }
 294 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
 295 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
 296 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
 297 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 298                     struct rds_sock *rs, u32 *key_ret);
 299 void rds_ib_sync_mr(void *trans_private, int dir);
 300 void rds_ib_free_mr(void *trans_private, int invalidate);
 301 void rds_ib_flush_mrs(void);
 302
 303 /* ib_recv.c */
 304 int __init rds_ib_recv_init(void);
 305 void rds_ib_recv_exit(void);
 306 int rds_ib_recv(struct rds_connection *conn);
 307 int rds_ib_recv_refill(struct rds_connection *conn, int prefill);
 308 void rds_ib_inc_free(struct rds_incoming *inc);
 309 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
 310                              size_t size);
 311 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
 312 void rds_ib_recv_tasklet_fn(unsigned long data);
 313 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
 314 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
 315 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
 316 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
 317 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
 318 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
 319
 320 /* ib_ring.c */
 321 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
 322 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
 323 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
 324 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
 325 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
 326 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
 327 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
 328 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
 329 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
 330 extern wait_queue_head_t rds_ib_ring_empty_wait;
 331
 332 /* ib_send.c */
 333 void rds_ib_xmit_complete(struct rds_connection *conn);
 334 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 335                 unsigned int hdr_off, unsigned int sg, unsigned int off);
 336 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
 337 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
 338 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
 339 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
 340 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
 341 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
 342 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
 343                              u32 *adv_credits, int need_posted, int max_posted);
 344 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
 345
 346 /* ib_stats.c */
 347 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
 348 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
 349 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
 350                                     unsigned int avail);
 351
 352 /* ib_sysctl.c */
 353 int __init rds_ib_sysctl_init(void);
 354 void rds_ib_sysctl_exit(void);
 355 extern unsigned long rds_ib_sysctl_max_send_wr;
 356 extern unsigned long rds_ib_sysctl_max_recv_wr;
 357 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
 358 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
 359 extern unsigned long rds_ib_sysctl_max_recv_allocation;
 360 extern unsigned int rds_ib_sysctl_flow_control;
 361 extern ctl_table rds_ib_sysctl_table[];
 362
 363 #endif