net/smc/smc_core.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /*
   3  * Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  Definitions for SMC Connections, Link Groups and Links
   6  *
   7  *  Copyright IBM Corp. 2016
   8  *
   9  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  10  */
  11
  12 #ifndef _SMC_CORE_H
  13 #define _SMC_CORE_H
  14
  15 #include <linux/atomic.h>
  16 #include <linux/smc.h>
  17 #include <linux/pci.h>
  18 #include <rdma/ib_verbs.h>
  19 #include <net/genetlink.h>
  20
  21 #include "smc.h"
  22 #include "smc_ib.h"
  23
  24 #define SMC_RMBS_PER_LGR_MAX    255     /* max. # of RMBs per link group */
  25
  26 struct smc_lgr_list {                   /* list of link group definition */
  27         struct list_head        list;
  28         spinlock_t              lock;   /* protects list of link groups */
  29         u32                     num;    /* unique link group number */
  30 };
  31
  32 enum smc_lgr_role {             /* possible roles of a link group */
  33         SMC_CLNT,       /* client */
  34         SMC_SERV        /* server */
  35 };
  36
  37 enum smc_link_state {                   /* possible states of a link */
  38         SMC_LNK_UNUSED,         /* link is unused */
  39         SMC_LNK_INACTIVE,       /* link is inactive */
  40         SMC_LNK_ACTIVATING,     /* link is being activated */
  41         SMC_LNK_ACTIVE,         /* link is active */
  42 };
  43
  44 #define SMC_WR_BUF_SIZE         48      /* size of work request buffer */
  45
  46 struct smc_wr_buf {
  47         u8      raw[SMC_WR_BUF_SIZE];
  48 };
  49
  50 #define SMC_WR_REG_MR_WAIT_TIME (5 * HZ)/* wait time for ib_wr_reg_mr result */
  51
  52 enum smc_wr_reg_state {
  53         POSTED,         /* ib_wr_reg_mr request posted */
  54         CONFIRMED,      /* ib_wr_reg_mr response: successful */
  55         FAILED          /* ib_wr_reg_mr response: failure */
  56 };
  57
  58 struct smc_rdma_sge {                           /* sges for RDMA writes */
  59         struct ib_sge           wr_tx_rdma_sge[SMC_IB_MAX_SEND_SGE];
  60 };
  61
  62 #define SMC_MAX_RDMA_WRITES     2               /* max. # of RDMA writes per
  63                                                  * message send
  64                                                  */
  65
  66 struct smc_rdma_sges {                          /* sges per message send */
  67         struct smc_rdma_sge     tx_rdma_sge[SMC_MAX_RDMA_WRITES];
  68 };
  69
  70 struct smc_rdma_wr {                            /* work requests per message
  71                                                  * send
  72                                                  */
  73         struct ib_rdma_wr       wr_tx_rdma[SMC_MAX_RDMA_WRITES];
  74 };
  75
  76 #define SMC_LGR_ID_SIZE         4
  77
  78 struct smc_link {
  79         struct smc_ib_device    *smcibdev;      /* ib-device */
  80         u8                      ibport;         /* port - values 1 | 2 */
  81         struct ib_pd            *roce_pd;       /* IB protection domain,
  82                                                  * unique for every RoCE QP
  83                                                  */
  84         struct ib_qp            *roce_qp;       /* IB queue pair */
  85         struct ib_qp_attr       qp_attr;        /* IB queue pair attributes */
  86
  87         struct smc_wr_buf       *wr_tx_bufs;    /* WR send payload buffers */
  88         struct ib_send_wr       *wr_tx_ibs;     /* WR send meta data */
  89         struct ib_sge           *wr_tx_sges;    /* WR send gather meta data */
  90         struct smc_rdma_sges    *wr_tx_rdma_sges;/*RDMA WRITE gather meta data*/
  91         struct smc_rdma_wr      *wr_tx_rdmas;   /* WR RDMA WRITE */
  92         struct smc_wr_tx_pend   *wr_tx_pends;   /* WR send waiting for CQE */
  93         struct completion       *wr_tx_compl;   /* WR send CQE completion */
  94         /* above four vectors have wr_tx_cnt elements and use the same index */
  95         dma_addr_t              wr_tx_dma_addr; /* DMA address of wr_tx_bufs */
  96         atomic_long_t           wr_tx_id;       /* seq # of last sent WR */
  97         unsigned long           *wr_tx_mask;    /* bit mask of used indexes */
  98         u32                     wr_tx_cnt;      /* number of WR send buffers */
  99         wait_queue_head_t       wr_tx_wait;     /* wait for free WR send buf */
 100
 101         struct smc_wr_buf       *wr_rx_bufs;    /* WR recv payload buffers */
 102         struct ib_recv_wr       *wr_rx_ibs;     /* WR recv meta data */
 103         struct ib_sge           *wr_rx_sges;    /* WR recv scatter meta data */
 104         /* above three vectors have wr_rx_cnt elements and use the same index */
 105         dma_addr_t              wr_rx_dma_addr; /* DMA address of wr_rx_bufs */
 106         u64                     wr_rx_id;       /* seq # of last recv WR */
 107         u32                     wr_rx_cnt;      /* number of WR recv buffers */
 108         unsigned long           wr_rx_tstamp;   /* jiffies when last buf rx */
 109
 110         struct ib_reg_wr        wr_reg;         /* WR register memory region */
 111         wait_queue_head_t       wr_reg_wait;    /* wait for wr_reg result */
 112         enum smc_wr_reg_state   wr_reg_state;   /* state of wr_reg request */
 113
 114         u8                      gid[SMC_GID_SIZE];/* gid matching used vlan id*/
 115         u8                      sgid_index;     /* gid index for vlan id      */
 116         u32                     peer_qpn;       /* QP number of peer */
 117         enum ib_mtu             path_mtu;       /* used mtu */
 118         enum ib_mtu             peer_mtu;       /* mtu size of peer */
 119         u32                     psn_initial;    /* QP tx initial packet seqno */
 120         u32                     peer_psn;       /* QP rx initial packet seqno */
 121         u8                      peer_mac[ETH_ALEN];     /* = gid[8:10||13:15] */
 122         u8                      peer_gid[SMC_GID_SIZE]; /* gid of peer*/
 123         u8                      link_id;        /* unique # within link group */
 124         u8                      link_uid[SMC_LGR_ID_SIZE]; /* unique lnk id */
 125         u8                      peer_link_uid[SMC_LGR_ID_SIZE]; /* peer uid */
 126         u8                      link_idx;       /* index in lgr link array */
 127         u8                      link_is_asym;   /* is link asymmetric? */
 128         struct smc_link_group   *lgr;           /* parent link group */
 129         struct work_struct      link_down_wrk;  /* wrk to bring link down */
 130         char                    ibname[IB_DEVICE_NAME_MAX]; /* ib device name */
 131         int                     ndev_ifidx; /* network device ifindex */
 132
 133         enum smc_link_state     state;          /* state of link */
 134         struct delayed_work     llc_testlink_wrk; /* testlink worker */
 135         struct completion       llc_testlink_resp; /* wait for rx of testlink */
 136         int                     llc_testlink_time; /* testlink interval */
 137         atomic_t                conn_cnt; /* connections on this link */
 138 };
 139
 140 /* For now we just allow one parallel link per link group. The SMC protocol
 141  * allows more (up to 8).
 142  */
 143 #define SMC_LINKS_PER_LGR_MAX   3
 144 #define SMC_SINGLE_LINK         0
 145
 146 /* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */
 147 struct smc_buf_desc {
 148         struct list_head        list;
 149         void                    *cpu_addr;      /* virtual address of buffer */
 150         struct page             *pages;
 151         int                     len;            /* length of buffer */
 152         u32                     used;           /* currently used / unused */
 153         union {
 154                 struct { /* SMC-R */
 155                         struct sg_table sgt[SMC_LINKS_PER_LGR_MAX];
 156                                         /* virtual buffer */
 157                         struct ib_mr    *mr_rx[SMC_LINKS_PER_LGR_MAX];
 158                                         /* for rmb only: memory region
 159                                          * incl. rkey provided to peer
 160                                          */
 161                         u32             order;  /* allocation order */
 162
 163                         u8              is_conf_rkey;
 164                                         /* confirm_rkey done */
 165                         u8              is_reg_mr[SMC_LINKS_PER_LGR_MAX];
 166                                         /* mem region registered */
 167                         u8              is_map_ib[SMC_LINKS_PER_LGR_MAX];
 168                                         /* mem region mapped to lnk */
 169                         u8              is_reg_err;
 170                                         /* buffer registration err */
 171                 };
 172                 struct { /* SMC-D */
 173                         unsigned short  sba_idx;
 174                                         /* SBA index number */
 175                         u64             token;
 176                                         /* DMB token number */
 177                         dma_addr_t      dma_addr;
 178                                         /* DMA address */
 179                 };
 180         };
 181 };
 182
 183 struct smc_rtoken {                             /* address/key of remote RMB */
 184         u64                     dma_addr;
 185         u32                     rkey;
 186 };
 187
 188 #define SMC_BUF_MIN_SIZE        16384   /* minimum size of an RMB */
 189 #define SMC_RMBE_SIZES          16      /* number of distinct RMBE sizes */
 190 /* theoretically, the RFC states that largest size would be 512K,
 191  * i.e. compressed 5 and thus 6 sizes (0..5), despite
 192  * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
 193  */
 194
 195 struct smcd_dev;
 196
 197 enum smc_lgr_type {                             /* redundancy state of lgr */
 198         SMC_LGR_NONE,                   /* no active links, lgr to be deleted */
 199         SMC_LGR_SINGLE,                 /* 1 active RNIC on each peer */
 200         SMC_LGR_SYMMETRIC,              /* 2 active RNICs on each peer */
 201         SMC_LGR_ASYMMETRIC_PEER,        /* local has 2, peer 1 active RNICs */
 202         SMC_LGR_ASYMMETRIC_LOCAL,       /* local has 1, peer 2 active RNICs */
 203 };
 204
 205 enum smc_llc_flowtype {
 206         SMC_LLC_FLOW_NONE       = 0,
 207         SMC_LLC_FLOW_ADD_LINK   = 2,
 208         SMC_LLC_FLOW_DEL_LINK   = 4,
 209         SMC_LLC_FLOW_RKEY       = 6,
 210 };
 211
 212 struct smc_llc_qentry;
 213
 214 struct smc_llc_flow {
 215         enum smc_llc_flowtype type;
 216         struct smc_llc_qentry *qentry;
 217 };
 218
 219 struct smc_link_group {
 220         struct list_head        list;
 221         struct rb_root          conns_all;      /* connection tree */
 222         rwlock_t                conns_lock;     /* protects conns_all */
 223         unsigned int            conns_num;      /* current # of connections */
 224         unsigned short          vlan_id;        /* vlan id of link group */
 225
 226         struct list_head        sndbufs[SMC_RMBE_SIZES];/* tx buffers */
 227         struct mutex            sndbufs_lock;   /* protects tx buffers */
 228         struct list_head        rmbs[SMC_RMBE_SIZES];   /* rx buffers */
 229         struct mutex            rmbs_lock;      /* protects rx buffers */
 230
 231         u8                      id[SMC_LGR_ID_SIZE];    /* unique lgr id */
 232         struct delayed_work     free_work;      /* delayed freeing of an lgr */
 233         struct work_struct      terminate_work; /* abnormal lgr termination */
 234         struct workqueue_struct *tx_wq;         /* wq for conn. tx workers */
 235         u8                      sync_err : 1;   /* lgr no longer fits to peer */
 236         u8                      terminating : 1;/* lgr is terminating */
 237         u8                      freeing : 1;    /* lgr is being freed */
 238
 239         bool                    is_smcd;        /* SMC-R or SMC-D */
 240         u8                      smc_version;
 241         u8                      negotiated_eid[SMC_MAX_EID_LEN];
 242         u8                      peer_os;        /* peer operating system */
 243         u8                      peer_smc_release;
 244         u8                      peer_hostname[SMC_MAX_HOSTNAME_LEN];
 245         union {
 246                 struct { /* SMC-R */
 247                         enum smc_lgr_role       role;
 248                                                 /* client or server */
 249                         struct smc_link         lnk[SMC_LINKS_PER_LGR_MAX];
 250                                                 /* smc link */
 251                         char                    peer_systemid[SMC_SYSTEMID_LEN];
 252                                                 /* unique system_id of peer */
 253                         struct smc_rtoken       rtokens[SMC_RMBS_PER_LGR_MAX]
 254                                                 [SMC_LINKS_PER_LGR_MAX];
 255                                                 /* remote addr/key pairs */
 256                         DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX);
 257                                                 /* used rtoken elements */
 258                         u8                      next_link_id;
 259                         enum smc_lgr_type       type;
 260                                                 /* redundancy state */
 261                         u8                      pnet_id[SMC_MAX_PNETID_LEN + 1];
 262                                                 /* pnet id of this lgr */
 263                         struct list_head        llc_event_q;
 264                                                 /* queue for llc events */
 265                         spinlock_t              llc_event_q_lock;
 266                                                 /* protects llc_event_q */
 267                         struct mutex            llc_conf_mutex;
 268                                                 /* protects lgr reconfig. */
 269                         struct work_struct      llc_add_link_work;
 270                         struct work_struct      llc_del_link_work;
 271                         struct work_struct      llc_event_work;
 272                                                 /* llc event worker */
 273                         wait_queue_head_t       llc_flow_waiter;
 274                                                 /* w4 next llc event */
 275                         wait_queue_head_t       llc_msg_waiter;
 276                                                 /* w4 next llc msg */
 277                         struct smc_llc_flow     llc_flow_lcl;
 278                                                 /* llc local control field */
 279                         struct smc_llc_flow     llc_flow_rmt;
 280                                                 /* llc remote control field */
 281                         struct smc_llc_qentry   *delayed_event;
 282                                                 /* arrived when flow active */
 283                         spinlock_t              llc_flow_lock;
 284                                                 /* protects llc flow */
 285                         int                     llc_testlink_time;
 286                                                 /* link keep alive time */
 287                         u32                     llc_termination_rsn;
 288                                                 /* rsn code for termination */
 289                 };
 290                 struct { /* SMC-D */
 291                         u64                     peer_gid;
 292                                                 /* Peer GID (remote) */
 293                         struct smcd_dev         *smcd;
 294                                                 /* ISM device for VLAN reg. */
 295                         u8                      peer_shutdown : 1;
 296                                                 /* peer triggered shutdownn */
 297                 };
 298         };
 299 };
 300
 301 struct smc_clc_msg_local;
 302
 303 struct smc_init_info {
 304         u8                      is_smcd;
 305         u8                      smc_type_v1;
 306         u8                      smc_type_v2;
 307         u8                      first_contact_peer;
 308         u8                      first_contact_local;
 309         unsigned short          vlan_id;
 310         u32                     rc;
 311         /* SMC-R */
 312         struct smc_clc_msg_local *ib_lcl;
 313         struct smc_ib_device    *ib_dev;
 314         u8                      ib_gid[SMC_GID_SIZE];
 315         u8                      ib_port;
 316         u32                     ib_clcqpn;
 317         /* SMC-D */
 318         u64                     ism_peer_gid[SMC_MAX_ISM_DEVS + 1];
 319         struct smcd_dev         *ism_dev[SMC_MAX_ISM_DEVS + 1];
 320         u16                     ism_chid[SMC_MAX_ISM_DEVS + 1];
 321         u8                      ism_offered_cnt; /* # of ISM devices offered */
 322         u8                      ism_selected;    /* index of selected ISM dev*/
 323         u8                      smcd_version;
 324 };
 325
 326 /* Find the connection associated with the given alert token in the link group.
 327  * To use rbtrees we have to implement our own search core.
 328  * Requires @conns_lock
 329  * @token       alert token to search for
 330  * @lgr          link group to search in
 331  * Returns connection associated with token if found, NULL otherwise.
 332  */
 333 static inline struct smc_connection *smc_lgr_find_conn(
 334         u32 token, struct smc_link_group *lgr)
 335 {
 336         struct smc_connection *res = NULL;
 337         struct rb_node *node;
 338
 339         node = lgr->conns_all.rb_node;
 340         while (node) {
 341                 struct smc_connection *cur = rb_entry(node,
 342                                         struct smc_connection, alert_node);
 343
 344                 if (cur->alert_token_local > token) {
 345                         node = node->rb_left;
 346                 } else {
 347                         if (cur->alert_token_local < token) {
 348                                 node = node->rb_right;
 349                         } else {
 350                                 res = cur;
 351                                 break;
 352                         }
 353                 }
 354         }
 355
 356         return res;
 357 }
 358
 359 /* returns true if the specified link is usable */
 360 static inline bool smc_link_usable(struct smc_link *lnk)
 361 {
 362         if (lnk->state == SMC_LNK_UNUSED || lnk->state == SMC_LNK_INACTIVE)
 363                 return false;
 364         return true;
 365 }
 366
 367 static inline bool smc_link_active(struct smc_link *lnk)
 368 {
 369         return lnk->state == SMC_LNK_ACTIVE;
 370 }
 371
 372 static inline void smc_gid_be16_convert(__u8 *buf, u8 *gid_raw)
 373 {
 374         sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x",
 375                 be16_to_cpu(((__be16 *)gid_raw)[0]),
 376                 be16_to_cpu(((__be16 *)gid_raw)[1]),
 377                 be16_to_cpu(((__be16 *)gid_raw)[2]),
 378                 be16_to_cpu(((__be16 *)gid_raw)[3]),
 379                 be16_to_cpu(((__be16 *)gid_raw)[4]),
 380                 be16_to_cpu(((__be16 *)gid_raw)[5]),
 381                 be16_to_cpu(((__be16 *)gid_raw)[6]),
 382                 be16_to_cpu(((__be16 *)gid_raw)[7]));
 383 }
 384
 385 struct smc_pci_dev {
 386         __u32           pci_fid;
 387         __u16           pci_pchid;
 388         __u16           pci_vendor;
 389         __u16           pci_device;
 390         __u8            pci_id[SMC_PCI_ID_STR_LEN];
 391 };
 392
 393 static inline void smc_set_pci_values(struct pci_dev *pci_dev,
 394                                       struct smc_pci_dev *smc_dev)
 395 {
 396         smc_dev->pci_vendor = pci_dev->vendor;
 397         smc_dev->pci_device = pci_dev->device;
 398         snprintf(smc_dev->pci_id, sizeof(smc_dev->pci_id), "%s",
 399                  pci_name(pci_dev));
 400 #if IS_ENABLED(CONFIG_S390)
 401         { /* Set s390 specific PCI information */
 402         struct zpci_dev *zdev;
 403
 404         zdev = to_zpci(pci_dev);
 405         smc_dev->pci_fid = zdev->fid;
 406         smc_dev->pci_pchid = zdev->pchid;
 407         }
 408 #endif
 409 }
 410
 411 struct smc_sock;
 412 struct smc_clc_msg_accept_confirm;
 413 struct smc_clc_msg_local;
 414
 415 void smc_lgr_cleanup_early(struct smc_connection *conn);
 416 void smc_lgr_terminate_sched(struct smc_link_group *lgr);
 417 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport);
 418 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport);
 419 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid,
 420                         unsigned short vlan);
 421 void smc_smcd_terminate_all(struct smcd_dev *dev);
 422 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev);
 423 int smc_buf_create(struct smc_sock *smc, bool is_smcd);
 424 int smc_uncompress_bufsize(u8 compressed);
 425 int smc_rmb_rtoken_handling(struct smc_connection *conn, struct smc_link *link,
 426                             struct smc_clc_msg_accept_confirm *clc);
 427 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey);
 428 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey);
 429 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
 430                     __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey);
 431 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
 432                      __be64 nw_vaddr, __be32 nw_rkey);
 433 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn);
 434 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn);
 435 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn);
 436 void smc_rmb_sync_sg_for_device(struct smc_connection *conn);
 437 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini);
 438
 439 void smc_conn_free(struct smc_connection *conn);
 440 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini);
 441 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr);
 442 int smc_core_init(void);
 443 void smc_core_exit(void);
 444
 445 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
 446                    u8 link_idx, struct smc_init_info *ini);
 447 void smcr_link_clear(struct smc_link *lnk, bool log);
 448 int smcr_buf_map_lgr(struct smc_link *lnk);
 449 int smcr_buf_reg_lgr(struct smc_link *lnk);
 450 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type);
 451 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
 452                             enum smc_lgr_type new_type, int asym_lnk_idx);
 453 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc);
 454 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
 455                                   struct smc_link *from_lnk, bool is_dev_err);
 456 void smcr_link_down_cond(struct smc_link *lnk);
 457 void smcr_link_down_cond_sched(struct smc_link *lnk);
 458 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb);
 459 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb);
 460 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb);
 461 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb);
 462
 463 static inline struct smc_link_group *smc_get_lgr(struct smc_link *link)
 464 {
 465         return link->lgr;
 466 }
 467 #endif