net/smc/smc_cdc.c

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  * Shared Memory Communications over RDMA (SMC-R) and RoCE
0004  *
0005  * Connection Data Control (CDC)
0006  * handles flow control
0007  *
0008  * Copyright IBM Corp. 2016
0009  *
0010  * Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
0011  */
0012
0013 #include <linux/spinlock.h>
0014
0015 #include "smc.h"
0016 #include "smc_wr.h"
0017 #include "smc_cdc.h"
0018 #include "smc_tx.h"
0019 #include "smc_rx.h"
0020 #include "smc_close.h"
0021
0022 /********************************** send *************************************/
0023
0024 /* handler for send/transmission completion of a CDC msg */
0025 static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd,
0026                    struct smc_link *link,
0027                    enum ib_wc_status wc_status)
0028 {
0029     struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd;
0030     struct smc_connection *conn = cdcpend->conn;
0031     struct smc_sock *smc;
0032     int diff;
0033
0034     smc = container_of(conn, struct smc_sock, conn);
0035     bh_lock_sock(&smc->sk);
0036     if (!wc_status) {
0037         diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len,
0038                      &cdcpend->conn->tx_curs_fin,
0039                      &cdcpend->cursor);
0040         /* sndbuf_space is decreased in smc_sendmsg */
0041         smp_mb__before_atomic();
0042         atomic_add(diff, &cdcpend->conn->sndbuf_space);
0043         /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
0044         smp_mb__after_atomic();
0045         smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn);
0046         smc_curs_copy(&conn->local_tx_ctrl_fin, &cdcpend->p_cursor,
0047                   conn);
0048         conn->tx_cdc_seq_fin = cdcpend->ctrl_seq;
0049     }
0050
0051     if (atomic_dec_and_test(&conn->cdc_pend_tx_wr)) {
0052         /* If user owns the sock_lock, mark the connection need sending.
0053          * User context will later try to send when it release sock_lock
0054          * in smc_release_cb()
0055          */
0056         if (sock_owned_by_user(&smc->sk))
0057             conn->tx_in_release_sock = true;
0058         else
0059             smc_tx_pending(conn);
0060
0061         if (unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq)))
0062             wake_up(&conn->cdc_pend_tx_wq);
0063     }
0064     WARN_ON(atomic_read(&conn->cdc_pend_tx_wr) < 0);
0065
0066     smc_tx_sndbuf_nonfull(smc);
0067     bh_unlock_sock(&smc->sk);
0068 }
0069
0070 int smc_cdc_get_free_slot(struct smc_connection *conn,
0071               struct smc_link *link,
0072               struct smc_wr_buf **wr_buf,
0073               struct smc_rdma_wr **wr_rdma_buf,
0074               struct smc_cdc_tx_pend **pend)
0075 {
0076     int rc;
0077
0078     rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
0079                      wr_rdma_buf,
0080                      (struct smc_wr_tx_pend_priv **)pend);
0081     if (conn->killed) {
0082         /* abnormal termination */
0083         if (!rc)
0084             smc_wr_tx_put_slot(link,
0085                        (struct smc_wr_tx_pend_priv *)(*pend));
0086         rc = -EPIPE;
0087     }
0088     return rc;
0089 }
0090
0091 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
0092                         struct smc_cdc_tx_pend *pend)
0093 {
0094     BUILD_BUG_ON_MSG(
0095         sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
0096         "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
0097     BUILD_BUG_ON_MSG(
0098         offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
0099         "must adapt SMC_WR_TX_SIZE to sizeof(struct smc_cdc_msg); if not all smc_wr upper layer protocols use the same message size any more, must start to set link->wr_tx_sges[i].length on each individual smc_wr_tx_send()");
0100     BUILD_BUG_ON_MSG(
0101         sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
0102         "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
0103     pend->conn = conn;
0104     pend->cursor = conn->tx_curs_sent;
0105     pend->p_cursor = conn->local_tx_ctrl.prod;
0106     pend->ctrl_seq = conn->tx_cdc_seq;
0107 }
0108
0109 int smc_cdc_msg_send(struct smc_connection *conn,
0110              struct smc_wr_buf *wr_buf,
0111              struct smc_cdc_tx_pend *pend)
0112 {
0113     struct smc_link *link = conn->lnk;
0114     union smc_host_cursor cfed;
0115     int rc;
0116
0117     smc_cdc_add_pending_send(conn, pend);
0118
0119     conn->tx_cdc_seq++;
0120     conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
0121     smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
0122
0123     atomic_inc(&conn->cdc_pend_tx_wr);
0124     smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */
0125
0126     rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
0127     if (!rc) {
0128         smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
0129         conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
0130     } else {
0131         conn->tx_cdc_seq--;
0132         conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
0133         atomic_dec(&conn->cdc_pend_tx_wr);
0134     }
0135
0136     return rc;
0137 }
0138
0139 /* send a validation msg indicating the move of a conn to an other QP link */
0140 int smcr_cdc_msg_send_validation(struct smc_connection *conn,
0141                  struct smc_cdc_tx_pend *pend,
0142                  struct smc_wr_buf *wr_buf)
0143 {
0144     struct smc_host_cdc_msg *local = &conn->local_tx_ctrl;
0145     struct smc_link *link = conn->lnk;
0146     struct smc_cdc_msg *peer;
0147     int rc;
0148
0149     peer = (struct smc_cdc_msg *)wr_buf;
0150     peer->common.type = local->common.type;
0151     peer->len = local->len;
0152     peer->seqno = htons(conn->tx_cdc_seq_fin); /* seqno last compl. tx */
0153     peer->token = htonl(local->token);
0154     peer->prod_flags.failover_validation = 1;
0155
0156     /* We need to set pend->conn here to make sure smc_cdc_tx_handler()
0157      * can handle properly
0158      */
0159     smc_cdc_add_pending_send(conn, pend);
0160
0161     atomic_inc(&conn->cdc_pend_tx_wr);
0162     smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */
0163
0164     rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
0165     if (unlikely(rc))
0166         atomic_dec(&conn->cdc_pend_tx_wr);
0167
0168     return rc;
0169 }
0170
0171 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
0172 {
0173     struct smc_cdc_tx_pend *pend;
0174     struct smc_wr_buf *wr_buf;
0175     struct smc_link *link;
0176     bool again = false;
0177     int rc;
0178
0179 again:
0180     link = conn->lnk;
0181     if (!smc_wr_tx_link_hold(link))
0182         return -ENOLINK;
0183     rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend);
0184     if (rc)
0185         goto put_out;
0186
0187     spin_lock_bh(&conn->send_lock);
0188     if (link != conn->lnk) {
0189         /* link of connection changed, try again one time*/
0190         spin_unlock_bh(&conn->send_lock);
0191         smc_wr_tx_put_slot(link,
0192                    (struct smc_wr_tx_pend_priv *)pend);
0193         smc_wr_tx_link_put(link);
0194         if (again)
0195             return -ENOLINK;
0196         again = true;
0197         goto again;
0198     }
0199     rc = smc_cdc_msg_send(conn, wr_buf, pend);
0200     spin_unlock_bh(&conn->send_lock);
0201 put_out:
0202     smc_wr_tx_link_put(link);
0203     return rc;
0204 }
0205
0206 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
0207 {
0208     int rc;
0209
0210     if (!smc_conn_lgr_valid(conn) ||
0211         (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
0212         return -EPIPE;
0213
0214     if (conn->lgr->is_smcd) {
0215         spin_lock_bh(&conn->send_lock);
0216         rc = smcd_cdc_msg_send(conn);
0217         spin_unlock_bh(&conn->send_lock);
0218     } else {
0219         rc = smcr_cdc_get_slot_and_msg_send(conn);
0220     }
0221
0222     return rc;
0223 }
0224
0225 void smc_cdc_wait_pend_tx_wr(struct smc_connection *conn)
0226 {
0227     wait_event(conn->cdc_pend_tx_wq, !atomic_read(&conn->cdc_pend_tx_wr));
0228 }
0229
0230 /* Send a SMC-D CDC header.
0231  * This increments the free space available in our send buffer.
0232  * Also update the confirmed receive buffer with what was sent to the peer.
0233  */
0234 int smcd_cdc_msg_send(struct smc_connection *conn)
0235 {
0236     struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
0237     union smc_host_cursor curs;
0238     struct smcd_cdc_msg cdc;
0239     int rc, diff;
0240
0241     memset(&cdc, 0, sizeof(cdc));
0242     cdc.common.type = SMC_CDC_MSG_TYPE;
0243     curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
0244     cdc.prod.wrap = curs.wrap;
0245     cdc.prod.count = curs.count;
0246     curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
0247     cdc.cons.wrap = curs.wrap;
0248     cdc.cons.count = curs.count;
0249     cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
0250     cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
0251     rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
0252     if (rc)
0253         return rc;
0254     smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
0255     conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
0256     /* Calculate transmitted data and increment free send buffer space */
0257     diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
0258                  &conn->tx_curs_sent);
0259     /* increased by confirmed number of bytes */
0260     smp_mb__before_atomic();
0261     atomic_add(diff, &conn->sndbuf_space);
0262     /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
0263     smp_mb__after_atomic();
0264     smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
0265
0266     smc_tx_sndbuf_nonfull(smc);
0267     return rc;
0268 }
0269
0270 /********************************* receive ***********************************/
0271
0272 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
0273 {
0274     return (s16)(seq1 - seq2) < 0;
0275 }
0276
0277 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
0278                         int *diff_prod)
0279 {
0280     struct smc_connection *conn = &smc->conn;
0281     char *base;
0282
0283     /* new data included urgent business */
0284     smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
0285     conn->urg_state = SMC_URG_VALID;
0286     if (!sock_flag(&smc->sk, SOCK_URGINLINE))
0287         /* we'll skip the urgent byte, so don't account for it */
0288         (*diff_prod)--;
0289     base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
0290     if (conn->urg_curs.count)
0291         conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
0292     else
0293         conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
0294     sk_send_sigurg(&smc->sk);
0295 }
0296
0297 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc,
0298                  struct smc_link *link)
0299 {
0300     struct smc_connection *conn = &smc->conn;
0301     u16 recv_seq = ntohs(cdc->seqno);
0302     s16 diff;
0303
0304     /* check that seqnum was seen before */
0305     diff = conn->local_rx_ctrl.seqno - recv_seq;
0306     if (diff < 0) { /* diff larger than 0x7fff */
0307         /* drop connection */
0308         conn->out_of_sync = 1;  /* prevent any further receives */
0309         spin_lock_bh(&conn->send_lock);
0310         conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
0311         conn->lnk = link;
0312         spin_unlock_bh(&conn->send_lock);
0313         sock_hold(&smc->sk); /* sock_put in abort_work */
0314         if (!queue_work(smc_close_wq, &conn->abort_work))
0315             sock_put(&smc->sk);
0316     }
0317 }
0318
0319 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
0320                     struct smc_cdc_msg *cdc)
0321 {
0322     union smc_host_cursor cons_old, prod_old;
0323     struct smc_connection *conn = &smc->conn;
0324     int diff_cons, diff_prod;
0325
0326     smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
0327     smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
0328     smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
0329
0330     diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
0331                   &conn->local_rx_ctrl.cons);
0332     if (diff_cons) {
0333         /* peer_rmbe_space is decreased during data transfer with RDMA
0334          * write
0335          */
0336         smp_mb__before_atomic();
0337         atomic_add(diff_cons, &conn->peer_rmbe_space);
0338         /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
0339         smp_mb__after_atomic();
0340     }
0341
0342     diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
0343                   &conn->local_rx_ctrl.prod);
0344     if (diff_prod) {
0345         if (conn->local_rx_ctrl.prod_flags.urg_data_present)
0346             smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
0347         /* bytes_to_rcv is decreased in smc_recvmsg */
0348         smp_mb__before_atomic();
0349         atomic_add(diff_prod, &conn->bytes_to_rcv);
0350         /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
0351         smp_mb__after_atomic();
0352         smc->sk.sk_data_ready(&smc->sk);
0353     } else {
0354         if (conn->local_rx_ctrl.prod_flags.write_blocked)
0355             smc->sk.sk_data_ready(&smc->sk);
0356         if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
0357             conn->urg_state = SMC_URG_NOTYET;
0358     }
0359
0360     /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
0361     if ((diff_cons && smc_tx_prepared_sends(conn)) ||
0362         conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
0363         conn->local_rx_ctrl.prod_flags.urg_data_pending) {
0364         if (!sock_owned_by_user(&smc->sk))
0365             smc_tx_pending(conn);
0366         else
0367             conn->tx_in_release_sock = true;
0368     }
0369
0370     if (diff_cons && conn->urg_tx_pend &&
0371         atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
0372         /* urg data confirmed by peer, indicate we're ready for more */
0373         conn->urg_tx_pend = false;
0374         smc->sk.sk_write_space(&smc->sk);
0375     }
0376
0377     if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
0378         smc->sk.sk_err = ECONNRESET;
0379         conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
0380     }
0381     if (smc_cdc_rxed_any_close_or_senddone(conn)) {
0382         smc->sk.sk_shutdown |= RCV_SHUTDOWN;
0383         if (smc->clcsock && smc->clcsock->sk)
0384             smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
0385         sock_set_flag(&smc->sk, SOCK_DONE);
0386         sock_hold(&smc->sk); /* sock_put in close_work */
0387         if (!queue_work(smc_close_wq, &conn->close_work))
0388             sock_put(&smc->sk);
0389     }
0390 }
0391
0392 /* called under tasklet context */
0393 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
0394 {
0395     sock_hold(&smc->sk);
0396     bh_lock_sock(&smc->sk);
0397     smc_cdc_msg_recv_action(smc, cdc);
0398     bh_unlock_sock(&smc->sk);
0399     sock_put(&smc->sk); /* no free sk in softirq-context */
0400 }
0401
0402 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
0403  * handler to indicate update in the DMBE.
0404  *
0405  * Context:
0406  * - tasklet context
0407  */
0408 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t)
0409 {
0410     struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet);
0411     struct smcd_cdc_msg *data_cdc;
0412     struct smcd_cdc_msg cdc;
0413     struct smc_sock *smc;
0414
0415     if (!conn || conn->killed)
0416         return;
0417
0418     data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
0419     smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
0420     smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
0421     smc = container_of(conn, struct smc_sock, conn);
0422     smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
0423 }
0424
0425 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
0426  * receiver side.
0427  */
0428 void smcd_cdc_rx_init(struct smc_connection *conn)
0429 {
0430     tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet);
0431 }
0432
0433 /***************************** init, exit, misc ******************************/
0434
0435 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
0436 {
0437     struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
0438     struct smc_cdc_msg *cdc = buf;
0439     struct smc_connection *conn;
0440     struct smc_link_group *lgr;
0441     struct smc_sock *smc;
0442
0443     if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
0444         return; /* short message */
0445     if (cdc->len != SMC_WR_TX_SIZE)
0446         return; /* invalid message */
0447
0448     /* lookup connection */
0449     lgr = smc_get_lgr(link);
0450     read_lock_bh(&lgr->conns_lock);
0451     conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
0452     read_unlock_bh(&lgr->conns_lock);
0453     if (!conn || conn->out_of_sync)
0454         return;
0455     smc = container_of(conn, struct smc_sock, conn);
0456
0457     if (cdc->prod_flags.failover_validation) {
0458         smc_cdc_msg_validate(smc, cdc, link);
0459         return;
0460     }
0461     if (smc_cdc_before(ntohs(cdc->seqno),
0462                conn->local_rx_ctrl.seqno))
0463         /* received seqno is old */
0464         return;
0465
0466     smc_cdc_msg_recv(smc, cdc);
0467 }
0468
0469 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
0470     {
0471         .handler    = smc_cdc_rx_handler,
0472         .type       = SMC_CDC_MSG_TYPE
0473     },
0474     {
0475         .handler    = NULL,
0476     }
0477 };
0478
0479 int __init smc_cdc_init(void)
0480 {
0481     struct smc_wr_rx_handler *handler;
0482     int rc = 0;
0483
0484     for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
0485         INIT_HLIST_NODE(&handler->list);
0486         rc = smc_wr_rx_register_handler(handler);
0487         if (rc)
0488             break;
0489     }
0490     return rc;
0491 }