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 /**
  23  * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
  24  * @xprt: transport associated with these backchannel resources
  25  * @reqs: number of concurrent incoming requests to expect
  26  *
  27  * Returns 0 on success; otherwise a negative errno
  28  */
  29 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
  30 {
  31         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
  32
  33         r_xprt->rx_buf.rb_bc_srv_max_requests = RPCRDMA_BACKWARD_WRS >> 1;
  34         trace_xprtrdma_cb_setup(r_xprt, reqs);
  35         return 0;
  36 }
  37
  38 /**
  39  * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
  40  * @xprt: transport
  41  *
  42  * Returns maximum size, in bytes, of a backchannel message
  43  */
  44 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
  45 {
  46         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
  47         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
  48         size_t maxmsg;
  49
  50         maxmsg = min_t(unsigned int, ep->rep_inline_send, ep->rep_inline_recv);
  51         maxmsg = min_t(unsigned int, maxmsg, PAGE_SIZE);
  52         return maxmsg - RPCRDMA_HDRLEN_MIN;
  53 }
  54
  55 unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *xprt)
  56 {
  57         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
  58
  59         return r_xprt->rx_buf.rb_bc_srv_max_requests;
  60 }
  61
  62 static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
  63 {
  64         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
  65         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
  66         __be32 *p;
  67
  68         rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
  69         xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
  70                         rdmab_data(req->rl_rdmabuf), rqst);
  71
  72         p = xdr_reserve_space(&req->rl_stream, 28);
  73         if (unlikely(!p))
  74                 return -EIO;
  75         *p++ = rqst->rq_xid;
  76         *p++ = rpcrdma_version;
  77         *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
  78         *p++ = rdma_msg;
  79         *p++ = xdr_zero;
  80         *p++ = xdr_zero;
  81         *p = xdr_zero;
  82
  83         if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
  84                                       &rqst->rq_snd_buf, rpcrdma_noch))
  85                 return -EIO;
  86
  87         trace_xprtrdma_cb_reply(rqst);
  88         return 0;
  89 }
  90
  91 /**
  92  * xprt_rdma_bc_send_reply - marshal and send a backchannel reply
  93  * @rqst: RPC rqst with a backchannel RPC reply in rq_snd_buf
  94  *
  95  * Caller holds the transport's write lock.
  96  *
  97  * Returns:
  98  *      %0 if the RPC message has been sent
  99  *      %-ENOTCONN if the caller should reconnect and call again
 100  *      %-EIO if a permanent error occurred and the request was not
 101  *              sent. Do not try to send this message again.
 102  */
 103 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst)
 104 {
 105         struct rpc_xprt *xprt = rqst->rq_xprt;
 106         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 107         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 108         int rc;
 109
 110         if (!xprt_connected(xprt))
 111                 return -ENOTCONN;
 112
 113         if (!xprt_request_get_cong(xprt, rqst))
 114                 return -EBADSLT;
 115
 116         rc = rpcrdma_bc_marshal_reply(rqst);
 117         if (rc < 0)
 118                 goto failed_marshal;
 119
 120         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 121                 goto drop_connection;
 122         return 0;
 123
 124 failed_marshal:
 125         if (rc != -ENOTCONN)
 126                 return rc;
 127 drop_connection:
 128         xprt_rdma_close(xprt);
 129         return -ENOTCONN;
 130 }
 131
 132 /**
 133  * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
 134  * @xprt: transport associated with these backchannel resources
 135  * @reqs: number of incoming requests to destroy; ignored
 136  */
 137 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
 138 {
 139         struct rpc_rqst *rqst, *tmp;
 140
 141         spin_lock(&xprt->bc_pa_lock);
 142         list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
 143                 list_del(&rqst->rq_bc_pa_list);
 144                 spin_unlock(&xprt->bc_pa_lock);
 145
 146                 rpcrdma_req_destroy(rpcr_to_rdmar(rqst));
 147
 148                 spin_lock(&xprt->bc_pa_lock);
 149         }
 150         spin_unlock(&xprt->bc_pa_lock);
 151 }
 152
 153 /**
 154  * xprt_rdma_bc_free_rqst - Release a backchannel rqst
 155  * @rqst: request to release
 156  */
 157 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
 158 {
 159         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 160         struct rpc_xprt *xprt = rqst->rq_xprt;
 161
 162         rpcrdma_recv_buffer_put(req->rl_reply);
 163         req->rl_reply = NULL;
 164
 165         spin_lock(&xprt->bc_pa_lock);
 166         list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
 167         spin_unlock(&xprt->bc_pa_lock);
 168         xprt_put(xprt);
 169 }
 170
 171 static struct rpc_rqst *rpcrdma_bc_rqst_get(struct rpcrdma_xprt *r_xprt)
 172 {
 173         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 174         struct rpcrdma_req *req;
 175         struct rpc_rqst *rqst;
 176         size_t size;
 177
 178         spin_lock(&xprt->bc_pa_lock);
 179         rqst = list_first_entry_or_null(&xprt->bc_pa_list, struct rpc_rqst,
 180                                         rq_bc_pa_list);
 181         if (!rqst)
 182                 goto create_req;
 183         list_del(&rqst->rq_bc_pa_list);
 184         spin_unlock(&xprt->bc_pa_lock);
 185         return rqst;
 186
 187 create_req:
 188         spin_unlock(&xprt->bc_pa_lock);
 189
 190         /* Set a limit to prevent a remote from overrunning our resources.
 191          */
 192         if (xprt->bc_alloc_count >= RPCRDMA_BACKWARD_WRS)
 193                 return NULL;
 194
 195         size = min_t(size_t, r_xprt->rx_ep.rep_inline_recv, PAGE_SIZE);
 196         req = rpcrdma_req_create(r_xprt, size, GFP_KERNEL);
 197         if (!req)
 198                 return NULL;
 199
 200         xprt->bc_alloc_count++;
 201         rqst = &req->rl_slot;
 202         rqst->rq_xprt = xprt;
 203         __set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
 204         xdr_buf_init(&rqst->rq_snd_buf, rdmab_data(req->rl_sendbuf), size);
 205         return rqst;
 206 }
 207
 208 /**
 209  * rpcrdma_bc_receive_call - Handle a backward direction call
 210  * @r_xprt: transport receiving the call
 211  * @rep: receive buffer containing the call
 212  *
 213  * Operational assumptions:
 214  *    o Backchannel credits are ignored, just as the NFS server
 215  *      forechannel currently does
 216  *    o The ULP manages a replay cache (eg, NFSv4.1 sessions).
 217  *      No replay detection is done at the transport level
 218  */
 219 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
 220                              struct rpcrdma_rep *rep)
 221 {
 222         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 223         struct svc_serv *bc_serv;
 224         struct rpcrdma_req *req;
 225         struct rpc_rqst *rqst;
 226         struct xdr_buf *buf;
 227         size_t size;
 228         __be32 *p;
 229
 230         p = xdr_inline_decode(&rep->rr_stream, 0);
 231         size = xdr_stream_remaining(&rep->rr_stream);
 232
 233 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
 234         pr_info("RPC:       %s: callback XID %08x, length=%u\n",
 235                 __func__, be32_to_cpup(p), size);
 236         pr_info("RPC:       %s: %*ph\n", __func__, size, p);
 237 #endif
 238
 239         rqst = rpcrdma_bc_rqst_get(r_xprt);
 240         if (!rqst)
 241                 goto out_overflow;
 242
 243         rqst->rq_reply_bytes_recvd = 0;
 244         rqst->rq_xid = *p;
 245
 246         rqst->rq_private_buf.len = size;
 247
 248         buf = &rqst->rq_rcv_buf;
 249         memset(buf, 0, sizeof(*buf));
 250         buf->head[0].iov_base = p;
 251         buf->head[0].iov_len = size;
 252         buf->len = size;
 253
 254         /* The receive buffer has to be hooked to the rpcrdma_req
 255          * so that it is not released while the req is pointing
 256          * to its buffer, and so that it can be reposted after
 257          * the Upper Layer is done decoding it.
 258          */
 259         req = rpcr_to_rdmar(rqst);
 260         req->rl_reply = rep;
 261         trace_xprtrdma_cb_call(rqst);
 262
 263         /* Queue rqst for ULP's callback service */
 264         bc_serv = xprt->bc_serv;
 265         xprt_get(xprt);
 266         spin_lock(&bc_serv->sv_cb_lock);
 267         list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
 268         spin_unlock(&bc_serv->sv_cb_lock);
 269
 270         wake_up(&bc_serv->sv_cb_waitq);
 271
 272         r_xprt->rx_stats.bcall_count++;
 273         return;
 274
 275 out_overflow:
 276         pr_warn("RPC/RDMA backchannel overflow\n");
 277         xprt_force_disconnect(xprt);
 278         /* This receive buffer gets reposted automatically
 279          * when the connection is re-established.
 280          */
 281         return;
 282 }