net/sunrpc/xprtrdma/backchannel.c

   1 /*
   2  * Copyright (c) 2015 Oracle.  All rights reserved.
   3  *
   4  * Support for backward direction RPCs on RPC/RDMA.
   5  */
   6
   7 #include <linux/module.h>
   8 #include <linux/sunrpc/xprt.h>
   9 #include <linux/sunrpc/svc.h>
  10 #include <linux/sunrpc/svc_xprt.h>
  11
  12 #include "xprt_rdma.h"
  13
  14 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  15 # define RPCDBG_FACILITY        RPCDBG_TRANS
  16 #endif
  17
  18 #undef RPCRDMA_BACKCHANNEL_DEBUG
  19
  20 static void rpcrdma_bc_free_rqst(struct rpcrdma_xprt *r_xprt,
  21                                  struct rpc_rqst *rqst)
  22 {
  23         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
  24         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
  25
  26         spin_lock(&buf->rb_reqslock);
  27         list_del(&req->rl_all);
  28         spin_unlock(&buf->rb_reqslock);
  29
  30         rpcrdma_destroy_req(&r_xprt->rx_ia, req);
  31
  32         kfree(rqst);
  33 }
  34
  35 static int rpcrdma_bc_setup_rqst(struct rpcrdma_xprt *r_xprt,
  36                                  struct rpc_rqst *rqst)
  37 {
  38         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
  39         struct rpcrdma_regbuf *rb;
  40         struct rpcrdma_req *req;
  41         size_t size;
  42
  43         req = rpcrdma_create_req(r_xprt);
  44         if (IS_ERR(req))
  45                 return PTR_ERR(req);
  46         req->rl_backchannel = true;
  47
  48         rb = rpcrdma_alloc_regbuf(ia, RPCRDMA_HDRBUF_SIZE, GFP_KERNEL);
  49         if (IS_ERR(rb))
  50                 goto out_fail;
  51         req->rl_rdmabuf = rb;
  52
  53         size = r_xprt->rx_data.inline_rsize;
  54         rb = rpcrdma_alloc_regbuf(ia, size, GFP_KERNEL);
  55         if (IS_ERR(rb))
  56                 goto out_fail;
  57         req->rl_sendbuf = rb;
  58         xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base, size);
  59         rpcrdma_set_xprtdata(rqst, req);
  60         return 0;
  61
  62 out_fail:
  63         rpcrdma_bc_free_rqst(r_xprt, rqst);
  64         return -ENOMEM;
  65 }
  66
  67 /* Allocate and add receive buffers to the rpcrdma_buffer's
  68  * existing list of rep's. These are released when the
  69  * transport is destroyed.
  70  */
  71 static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt,
  72                                  unsigned int count)
  73 {
  74         struct rpcrdma_rep *rep;
  75         int rc = 0;
  76
  77         while (count--) {
  78                 rep = rpcrdma_create_rep(r_xprt);
  79                 if (IS_ERR(rep)) {
  80                         pr_err("RPC:       %s: reply buffer alloc failed\n",
  81                                __func__);
  82                         rc = PTR_ERR(rep);
  83                         break;
  84                 }
  85
  86                 rpcrdma_recv_buffer_put(rep);
  87         }
  88
  89         return rc;
  90 }
  91
  92 /**
  93  * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
  94  * @xprt: transport associated with these backchannel resources
  95  * @reqs: number of concurrent incoming requests to expect
  96  *
  97  * Returns 0 on success; otherwise a negative errno
  98  */
  99 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
 100 {
 101         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 102         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
 103         struct rpc_rqst *rqst;
 104         unsigned int i;
 105         int rc;
 106
 107         /* The backchannel reply path returns each rpc_rqst to the
 108          * bc_pa_list _after_ the reply is sent. If the server is
 109          * faster than the client, it can send another backward
 110          * direction request before the rpc_rqst is returned to the
 111          * list. The client rejects the request in this case.
 112          *
 113          * Twice as many rpc_rqsts are prepared to ensure there is
 114          * always an rpc_rqst available as soon as a reply is sent.
 115          */
 116         if (reqs > RPCRDMA_BACKWARD_WRS >> 1)
 117                 goto out_err;
 118
 119         for (i = 0; i < (reqs << 1); i++) {
 120                 rqst = kzalloc(sizeof(*rqst), GFP_KERNEL);
 121                 if (!rqst) {
 122                         pr_err("RPC:       %s: Failed to create bc rpc_rqst\n",
 123                                __func__);
 124                         goto out_free;
 125                 }
 126                 dprintk("RPC:       %s: new rqst %p\n", __func__, rqst);
 127
 128                 rqst->rq_xprt = &r_xprt->rx_xprt;
 129                 INIT_LIST_HEAD(&rqst->rq_list);
 130                 INIT_LIST_HEAD(&rqst->rq_bc_list);
 131
 132                 if (rpcrdma_bc_setup_rqst(r_xprt, rqst))
 133                         goto out_free;
 134
 135                 spin_lock_bh(&xprt->bc_pa_lock);
 136                 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
 137                 spin_unlock_bh(&xprt->bc_pa_lock);
 138         }
 139
 140         rc = rpcrdma_bc_setup_reps(r_xprt, reqs);
 141         if (rc)
 142                 goto out_free;
 143
 144         rc = rpcrdma_ep_post_extra_recv(r_xprt, reqs);
 145         if (rc)
 146                 goto out_free;
 147
 148         buffer->rb_bc_srv_max_requests = reqs;
 149         request_module("svcrdma");
 150
 151         return 0;
 152
 153 out_free:
 154         xprt_rdma_bc_destroy(xprt, reqs);
 155
 156 out_err:
 157         pr_err("RPC:       %s: setup backchannel transport failed\n", __func__);
 158         return -ENOMEM;
 159 }
 160
 161 /**
 162  * xprt_rdma_bc_up - Create transport endpoint for backchannel service
 163  * @serv: server endpoint
 164  * @net: network namespace
 165  *
 166  * The "xprt" is an implied argument: it supplies the name of the
 167  * backchannel transport class.
 168  *
 169  * Returns zero on success, negative errno on failure
 170  */
 171 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net)
 172 {
 173         int ret;
 174
 175         ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0);
 176         if (ret < 0)
 177                 return ret;
 178         return 0;
 179 }
 180
 181 /**
 182  * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
 183  * @xprt: transport
 184  *
 185  * Returns maximum size, in bytes, of a backchannel message
 186  */
 187 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
 188 {
 189         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 190         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 191         size_t maxmsg;
 192
 193         maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize);
 194         return maxmsg - RPCRDMA_HDRLEN_MIN;
 195 }
 196
 197 /**
 198  * rpcrdma_bc_marshal_reply - Send backwards direction reply
 199  * @rqst: buffer containing RPC reply data
 200  *
 201  * Returns zero on success.
 202  */
 203 int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
 204 {
 205         struct rpc_xprt *xprt = rqst->rq_xprt;
 206         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 207         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 208         struct rpcrdma_msg *headerp;
 209         size_t rpclen;
 210
 211         headerp = rdmab_to_msg(req->rl_rdmabuf);
 212         headerp->rm_xid = rqst->rq_xid;
 213         headerp->rm_vers = rpcrdma_version;
 214         headerp->rm_credit =
 215                         cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
 216         headerp->rm_type = rdma_msg;
 217         headerp->rm_body.rm_chunks[0] = xdr_zero;
 218         headerp->rm_body.rm_chunks[1] = xdr_zero;
 219         headerp->rm_body.rm_chunks[2] = xdr_zero;
 220
 221         rpclen = rqst->rq_svec[0].iov_len;
 222
 223 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
 224         pr_info("RPC:       %s: rpclen %zd headerp 0x%p lkey 0x%x\n",
 225                 __func__, rpclen, headerp, rdmab_lkey(req->rl_rdmabuf));
 226         pr_info("RPC:       %s: RPC/RDMA: %*ph\n",
 227                 __func__, (int)RPCRDMA_HDRLEN_MIN, headerp);
 228         pr_info("RPC:       %s:      RPC: %*ph\n",
 229                 __func__, (int)rpclen, rqst->rq_svec[0].iov_base);
 230 #endif
 231
 232         req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
 233         req->rl_send_iov[0].length = RPCRDMA_HDRLEN_MIN;
 234         req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf);
 235
 236         req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf);
 237         req->rl_send_iov[1].length = rpclen;
 238         req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf);
 239
 240         req->rl_niovs = 2;
 241         return 0;
 242 }
 243
 244 /**
 245  * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
 246  * @xprt: transport associated with these backchannel resources
 247  * @reqs: number of incoming requests to destroy; ignored
 248  */
 249 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
 250 {
 251         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 252         struct rpc_rqst *rqst, *tmp;
 253
 254         spin_lock_bh(&xprt->bc_pa_lock);
 255         list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
 256                 list_del(&rqst->rq_bc_pa_list);
 257                 spin_unlock_bh(&xprt->bc_pa_lock);
 258
 259                 rpcrdma_bc_free_rqst(r_xprt, rqst);
 260
 261                 spin_lock_bh(&xprt->bc_pa_lock);
 262         }
 263         spin_unlock_bh(&xprt->bc_pa_lock);
 264 }
 265
 266 /**
 267  * xprt_rdma_bc_free_rqst - Release a backchannel rqst
 268  * @rqst: request to release
 269  */
 270 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
 271 {
 272         struct rpc_xprt *xprt = rqst->rq_xprt;
 273
 274         dprintk("RPC:       %s: freeing rqst %p (req %p)\n",
 275                 __func__, rqst, rpcr_to_rdmar(rqst));
 276
 277         smp_mb__before_atomic();
 278         WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state));
 279         clear_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
 280         smp_mb__after_atomic();
 281
 282         spin_lock_bh(&xprt->bc_pa_lock);
 283         list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
 284         spin_unlock_bh(&xprt->bc_pa_lock);
 285 }
 286
 287 /**
 288  * rpcrdma_bc_receive_call - Handle a backward direction call
 289  * @xprt: transport receiving the call
 290  * @rep: receive buffer containing the call
 291  *
 292  * Called in the RPC reply handler, which runs in a tasklet.
 293  * Be quick about it.
 294  *
 295  * Operational assumptions:
 296  *    o Backchannel credits are ignored, just as the NFS server
 297  *      forechannel currently does
 298  *    o The ULP manages a replay cache (eg, NFSv4.1 sessions).
 299  *      No replay detection is done at the transport level
 300  */
 301 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
 302                              struct rpcrdma_rep *rep)
 303 {
 304         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 305         struct rpcrdma_msg *headerp;
 306         struct svc_serv *bc_serv;
 307         struct rpcrdma_req *req;
 308         struct rpc_rqst *rqst;
 309         struct xdr_buf *buf;
 310         size_t size;
 311         __be32 *p;
 312
 313         headerp = rdmab_to_msg(rep->rr_rdmabuf);
 314 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
 315         pr_info("RPC:       %s: callback XID %08x, length=%u\n",
 316                 __func__, be32_to_cpu(headerp->rm_xid), rep->rr_len);
 317         pr_info("RPC:       %s: %*ph\n", __func__, rep->rr_len, headerp);
 318 #endif
 319
 320         /* Sanity check:
 321          * Need at least enough bytes for RPC/RDMA header, as code
 322          * here references the header fields by array offset. Also,
 323          * backward calls are always inline, so ensure there
 324          * are some bytes beyond the RPC/RDMA header.
 325          */
 326         if (rep->rr_len < RPCRDMA_HDRLEN_MIN + 24)
 327                 goto out_short;
 328         p = (__be32 *)((unsigned char *)headerp + RPCRDMA_HDRLEN_MIN);
 329         size = rep->rr_len - RPCRDMA_HDRLEN_MIN;
 330
 331         /* Grab a free bc rqst */
 332         spin_lock(&xprt->bc_pa_lock);
 333         if (list_empty(&xprt->bc_pa_list)) {
 334                 spin_unlock(&xprt->bc_pa_lock);
 335                 goto out_overflow;
 336         }
 337         rqst = list_first_entry(&xprt->bc_pa_list,
 338                                 struct rpc_rqst, rq_bc_pa_list);
 339         list_del(&rqst->rq_bc_pa_list);
 340         spin_unlock(&xprt->bc_pa_lock);
 341         dprintk("RPC:       %s: using rqst %p\n", __func__, rqst);
 342
 343         /* Prepare rqst */
 344         rqst->rq_reply_bytes_recvd = 0;
 345         rqst->rq_bytes_sent = 0;
 346         rqst->rq_xid = headerp->rm_xid;
 347
 348         rqst->rq_private_buf.len = size;
 349         set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
 350
 351         buf = &rqst->rq_rcv_buf;
 352         memset(buf, 0, sizeof(*buf));
 353         buf->head[0].iov_base = p;
 354         buf->head[0].iov_len = size;
 355         buf->len = size;
 356
 357         /* The receive buffer has to be hooked to the rpcrdma_req
 358          * so that it can be reposted after the server is done
 359          * parsing it but just before sending the backward
 360          * direction reply.
 361          */
 362         req = rpcr_to_rdmar(rqst);
 363         dprintk("RPC:       %s: attaching rep %p to req %p\n",
 364                 __func__, rep, req);
 365         req->rl_reply = rep;
 366
 367         /* Defeat the retransmit detection logic in send_request */
 368         req->rl_connect_cookie = 0;
 369
 370         /* Queue rqst for ULP's callback service */
 371         bc_serv = xprt->bc_serv;
 372         spin_lock(&bc_serv->sv_cb_lock);
 373         list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
 374         spin_unlock(&bc_serv->sv_cb_lock);
 375
 376         wake_up(&bc_serv->sv_cb_waitq);
 377
 378         r_xprt->rx_stats.bcall_count++;
 379         return;
 380
 381 out_overflow:
 382         pr_warn("RPC/RDMA backchannel overflow\n");
 383         xprt_disconnect_done(xprt);
 384         /* This receive buffer gets reposted automatically
 385          * when the connection is re-established.
 386          */
 387         return;
 388
 389 out_short:
 390         pr_warn("RPC/RDMA short backward direction call\n");
 391
 392         if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
 393                 xprt_disconnect_done(xprt);
 394         else
 395                 pr_warn("RPC:       %s: reposting rep %p\n",
 396                         __func__, rep);
 397 }