net/sunrpc/xprtrdma/backchannel.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2015 Oracle.  All rights reserved.
   4  *
   5  * Support for backward direction RPCs on RPC/RDMA.
   6  */
   7
   8 #include <linux/sunrpc/xprt.h>
   9 #include <linux/sunrpc/svc.h>
  10 #include <linux/sunrpc/svc_xprt.h>
  11 #include <linux/sunrpc/svc_rdma.h>
  12
  13 #include "xprt_rdma.h"
  14 #include <trace/events/rpcrdma.h>
  15
  16 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  17 # define RPCDBG_FACILITY        RPCDBG_TRANS
  18 #endif
  19
  20 #undef RPCRDMA_BACKCHANNEL_DEBUG
  21
  22 static int rpcrdma_bc_setup_reqs(struct rpcrdma_xprt *r_xprt,
  23                                  unsigned int count)
  24 {
  25         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
  26         struct rpcrdma_req *req;
  27         struct rpc_rqst *rqst;
  28         unsigned int i;
  29
  30         for (i = 0; i < (count << 1); i++) {
  31                 struct rpcrdma_regbuf *rb;
  32                 size_t size;
  33
  34                 req = rpcrdma_create_req(r_xprt);
  35                 if (IS_ERR(req))
  36                         return PTR_ERR(req);
  37                 rqst = &req->rl_slot;
  38
  39                 rqst->rq_xprt = xprt;
  40                 INIT_LIST_HEAD(&rqst->rq_bc_list);
  41                 __set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
  42                 spin_lock(&xprt->bc_pa_lock);
  43                 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
  44                 spin_unlock(&xprt->bc_pa_lock);
  45
  46                 size = r_xprt->rx_data.inline_rsize;
  47                 rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL);
  48                 if (IS_ERR(rb))
  49                         goto out_fail;
  50                 req->rl_sendbuf = rb;
  51                 xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base,
  52                              min_t(size_t, size, PAGE_SIZE));
  53         }
  54         return 0;
  55
  56 out_fail:
  57         rpcrdma_req_destroy(req);
  58         return -ENOMEM;
  59 }
  60
  61 /**
  62  * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
  63  * @xprt: transport associated with these backchannel resources
  64  * @reqs: number of concurrent incoming requests to expect
  65  *
  66  * Returns 0 on success; otherwise a negative errno
  67  */
  68 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
  69 {
  70         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
  71         int rc;
  72
  73         /* The backchannel reply path returns each rpc_rqst to the
  74          * bc_pa_list _after_ the reply is sent. If the server is
  75          * faster than the client, it can send another backward
  76          * direction request before the rpc_rqst is returned to the
  77          * list. The client rejects the request in this case.
  78          *
  79          * Twice as many rpc_rqsts are prepared to ensure there is
  80          * always an rpc_rqst available as soon as a reply is sent.
  81          */
  82         if (reqs > RPCRDMA_BACKWARD_WRS >> 1)
  83                 goto out_err;
  84
  85         rc = rpcrdma_bc_setup_reqs(r_xprt, reqs);
  86         if (rc)
  87                 goto out_free;
  88
  89         r_xprt->rx_buf.rb_bc_srv_max_requests = reqs;
  90         trace_xprtrdma_cb_setup(r_xprt, reqs);
  91         return 0;
  92
  93 out_free:
  94         xprt_rdma_bc_destroy(xprt, reqs);
  95
  96 out_err:
  97         pr_err("RPC:       %s: setup backchannel transport failed\n", __func__);
  98         return -ENOMEM;
  99 }
 100
 101 /**
 102  * xprt_rdma_bc_up - Create transport endpoint for backchannel service
 103  * @serv: server endpoint
 104  * @net: network namespace
 105  *
 106  * The "xprt" is an implied argument: it supplies the name of the
 107  * backchannel transport class.
 108  *
 109  * Returns zero on success, negative errno on failure
 110  */
 111 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net)
 112 {
 113         int ret;
 114
 115         ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0);
 116         if (ret < 0)
 117                 return ret;
 118         return 0;
 119 }
 120
 121 /**
 122  * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
 123  * @xprt: transport
 124  *
 125  * Returns maximum size, in bytes, of a backchannel message
 126  */
 127 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
 128 {
 129         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 130         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 131         size_t maxmsg;
 132
 133         maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize);
 134         maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE);
 135         return maxmsg - RPCRDMA_HDRLEN_MIN;
 136 }
 137
 138 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
 139 {
 140         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 141         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 142         __be32 *p;
 143
 144         rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
 145         xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
 146                         req->rl_rdmabuf->rg_base);
 147
 148         p = xdr_reserve_space(&req->rl_stream, 28);
 149         if (unlikely(!p))
 150                 return -EIO;
 151         *p++ = rqst->rq_xid;
 152         *p++ = rpcrdma_version;
 153         *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
 154         *p++ = rdma_msg;
 155         *p++ = xdr_zero;
 156         *p++ = xdr_zero;
 157         *p = xdr_zero;
 158
 159         if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
 160                                       &rqst->rq_snd_buf, rpcrdma_noch))
 161                 return -EIO;
 162
 163         trace_xprtrdma_cb_reply(rqst);
 164         return 0;
 165 }
 166
 167 /**
 168  * xprt_rdma_bc_send_reply - marshal and send a backchannel reply
 169  * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf
 170  *
 171  * Caller holds the transport's write lock.
 172  *
 173  * Returns:
 174  *      %0 if the RPC message has been sent
 175  *      %-ENOTCONN if the caller should reconnect and call again
 176  *      %-EIO if a permanent error occurred and the request was not
 177  *              sent. Do not try to send this message again.
 178  */
 179 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst)
 180 {
 181         struct rpc_xprt *xprt = rqst->rq_xprt;
 182         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 183         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 184         int rc;
 185
 186         if (!xprt_connected(xprt))
 187                 return -ENOTCONN;
 188
 189         if (!xprt_request_get_cong(xprt, rqst))
 190                 return -EBADSLT;
 191
 192         rc = rpcrdma_bc_marshal_reply(rqst);
 193         if (rc < 0)
 194                 goto failed_marshal;
 195
 196         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 197                 goto drop_connection;
 198         return 0;
 199
 200 failed_marshal:
 201         if (rc != -ENOTCONN)
 202                 return rc;
 203 drop_connection:
 204         xprt_rdma_close(xprt);
 205         return -ENOTCONN;
 206 }
 207
 208 /**
 209  * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
 210  * @xprt: transport associated with these backchannel resources
 211  * @reqs: number of incoming requests to destroy; ignored
 212  */
 213 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
 214 {
 215         struct rpc_rqst *rqst, *tmp;
 216
 217         spin_lock(&xprt->bc_pa_lock);
 218         list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
 219                 list_del(&rqst->rq_bc_pa_list);
 220                 spin_unlock(&xprt->bc_pa_lock);
 221
 222                 rpcrdma_req_destroy(rpcr_to_rdmar(rqst));
 223
 224                 spin_lock(&xprt->bc_pa_lock);
 225         }
 226         spin_unlock(&xprt->bc_pa_lock);
 227 }
 228
 229 /**
 230  * xprt_rdma_bc_free_rqst - Release a backchannel rqst
 231  * @rqst: request to release
 232  */
 233 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
 234 {
 235         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 236         struct rpc_xprt *xprt = rqst->rq_xprt;
 237
 238         rpcrdma_recv_buffer_put(req->rl_reply);
 239         req->rl_reply = NULL;
 240
 241         spin_lock(&xprt->bc_pa_lock);
 242         list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
 243         spin_unlock(&xprt->bc_pa_lock);
 244 }
 245
 246 /**
 247  * rpcrdma_bc_receive_call - Handle a backward direction call
 248  * @r_xprt: transport receiving the call
 249  * @rep: receive buffer containing the call
 250  *
 251  * Operational assumptions:
 252  *    o Backchannel credits are ignored, just as the NFS server
 253  *      forechannel currently does
 254  *    o The ULP manages a replay cache (eg, NFSv4.1 sessions).
 255  *      No replay detection is done at the transport level
 256  */
 257 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
 258                              struct rpcrdma_rep *rep)
 259 {
 260         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 261         struct svc_serv *bc_serv;
 262         struct rpcrdma_req *req;
 263         struct rpc_rqst *rqst;
 264         struct xdr_buf *buf;
 265         size_t size;
 266         __be32 *p;
 267
 268         p = xdr_inline_decode(&rep->rr_stream, 0);
 269         size = xdr_stream_remaining(&rep->rr_stream);
 270
 271 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
 272         pr_info("RPC:       %s: callback XID %08x, length=%u\n",
 273                 __func__, be32_to_cpup(p), size);
 274         pr_info("RPC:       %s: %*ph\n", __func__, size, p);
 275 #endif
 276
 277         /* Grab a free bc rqst */
 278         spin_lock(&xprt->bc_pa_lock);
 279         if (list_empty(&xprt->bc_pa_list)) {
 280                 spin_unlock(&xprt->bc_pa_lock);
 281                 goto out_overflow;
 282         }
 283         rqst = list_first_entry(&xprt->bc_pa_list,
 284                                 struct rpc_rqst, rq_bc_pa_list);
 285         list_del(&rqst->rq_bc_pa_list);
 286         spin_unlock(&xprt->bc_pa_lock);
 287
 288         /* Prepare rqst */
 289         rqst->rq_reply_bytes_recvd = 0;
 290         rqst->rq_bytes_sent = 0;
 291         rqst->rq_xid = *p;
 292
 293         rqst->rq_private_buf.len = size;
 294
 295         buf = &rqst->rq_rcv_buf;
 296         memset(buf, 0, sizeof(*buf));
 297         buf->head[0].iov_base = p;
 298         buf->head[0].iov_len = size;
 299         buf->len = size;
 300
 301         /* The receive buffer has to be hooked to the rpcrdma_req
 302          * so that it is not released while the req is pointing
 303          * to its buffer, and so that it can be reposted after
 304          * the Upper Layer is done decoding it.
 305          */
 306         req = rpcr_to_rdmar(rqst);
 307         req->rl_reply = rep;
 308         trace_xprtrdma_cb_call(rqst);
 309
 310         /* Queue rqst for ULP's callback service */
 311         bc_serv = xprt->bc_serv;
 312         spin_lock(&bc_serv->sv_cb_lock);
 313         list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
 314         spin_unlock(&bc_serv->sv_cb_lock);
 315
 316         wake_up(&bc_serv->sv_cb_waitq);
 317
 318         r_xprt->rx_stats.bcall_count++;
 319         return;
 320
 321 out_overflow:
 322         pr_warn("RPC/RDMA backchannel overflow\n");
 323         xprt_force_disconnect(xprt);
 324         /* This receive buffer gets reposted automatically
 325          * when the connection is re-established.
 326          */
 327         return;
 328 }