2 * Copyright (c) 2015 Oracle. All rights reserved.
4 * Support for backward direction RPCs on RPC/RDMA.
7 #include <linux/module.h>
8 #include <linux/sunrpc/xprt.h>
9 #include <linux/sunrpc/svc.h>
10 #include <linux/sunrpc/svc_xprt.h>
12 #include "xprt_rdma.h"
14 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
15 # define RPCDBG_FACILITY RPCDBG_TRANS
18 #undef RPCRDMA_BACKCHANNEL_DEBUG
20 static void rpcrdma_bc_free_rqst(struct rpcrdma_xprt *r_xprt,
21 struct rpc_rqst *rqst)
23 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
24 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
26 spin_lock(&buf->rb_reqslock);
27 list_del(&req->rl_all);
28 spin_unlock(&buf->rb_reqslock);
30 rpcrdma_destroy_req(&r_xprt->rx_ia, req);
35 static int rpcrdma_bc_setup_rqst(struct rpcrdma_xprt *r_xprt,
36 struct rpc_rqst *rqst)
38 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
39 struct rpcrdma_regbuf *rb;
40 struct rpcrdma_req *req;
43 req = rpcrdma_create_req(r_xprt);
46 req->rl_backchannel = true;
48 size = r_xprt->rx_data.inline_wsize;
49 rb = rpcrdma_alloc_regbuf(ia, size, GFP_KERNEL);
54 size += r_xprt->rx_data.inline_rsize;
55 rb = rpcrdma_alloc_regbuf(ia, size, GFP_KERNEL);
60 /* so that rpcr_to_rdmar works when receiving a request */
61 rqst->rq_buffer = (void *)req->rl_sendbuf->rg_base;
62 xdr_buf_init(&rqst->rq_snd_buf, rqst->rq_buffer, size);
66 rpcrdma_bc_free_rqst(r_xprt, rqst);
70 /* Allocate and add receive buffers to the rpcrdma_buffer's
71 * existing list of rep's. These are released when the
72 * transport is destroyed.
74 static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt,
77 struct rpcrdma_rep *rep;
81 rep = rpcrdma_create_rep(r_xprt);
83 pr_err("RPC: %s: reply buffer alloc failed\n",
89 rpcrdma_recv_buffer_put(rep);
96 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
97 * @xprt: transport associated with these backchannel resources
98 * @reqs: number of concurrent incoming requests to expect
100 * Returns 0 on success; otherwise a negative errno
102 int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
104 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
105 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
106 struct rpc_rqst *rqst;
110 /* The backchannel reply path returns each rpc_rqst to the
111 * bc_pa_list _after_ the reply is sent. If the server is
112 * faster than the client, it can send another backward
113 * direction request before the rpc_rqst is returned to the
114 * list. The client rejects the request in this case.
116 * Twice as many rpc_rqsts are prepared to ensure there is
117 * always an rpc_rqst available as soon as a reply is sent.
119 if (reqs > RPCRDMA_BACKWARD_WRS >> 1)
122 for (i = 0; i < (reqs << 1); i++) {
123 rqst = kzalloc(sizeof(*rqst), GFP_KERNEL);
125 pr_err("RPC: %s: Failed to create bc rpc_rqst\n",
129 dprintk("RPC: %s: new rqst %p\n", __func__, rqst);
131 rqst->rq_xprt = &r_xprt->rx_xprt;
132 INIT_LIST_HEAD(&rqst->rq_list);
133 INIT_LIST_HEAD(&rqst->rq_bc_list);
135 if (rpcrdma_bc_setup_rqst(r_xprt, rqst))
138 spin_lock_bh(&xprt->bc_pa_lock);
139 list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
140 spin_unlock_bh(&xprt->bc_pa_lock);
143 rc = rpcrdma_bc_setup_reps(r_xprt, reqs);
147 rc = rpcrdma_ep_post_extra_recv(r_xprt, reqs);
151 buffer->rb_bc_srv_max_requests = reqs;
152 request_module("svcrdma");
157 xprt_rdma_bc_destroy(xprt, reqs);
160 pr_err("RPC: %s: setup backchannel transport failed\n", __func__);
165 * xprt_rdma_bc_up - Create transport endpoint for backchannel service
166 * @serv: server endpoint
167 * @net: network namespace
169 * The "xprt" is an implied argument: it supplies the name of the
170 * backchannel transport class.
172 * Returns zero on success, negative errno on failure
174 int xprt_rdma_bc_up(struct svc_serv *serv, struct net *net)
178 ret = svc_create_xprt(serv, "rdma-bc", net, PF_INET, 0, 0);
185 * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
188 * Returns maximum size, in bytes, of a backchannel message
190 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *xprt)
192 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
193 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
196 maxmsg = min_t(unsigned int, cdata->inline_rsize, cdata->inline_wsize);
197 return maxmsg - RPCRDMA_HDRLEN_MIN;
201 * rpcrdma_bc_marshal_reply - Send backwards direction reply
202 * @rqst: buffer containing RPC reply data
204 * Returns zero on success.
206 int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
208 struct rpc_xprt *xprt = rqst->rq_xprt;
209 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
210 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
211 struct rpcrdma_msg *headerp;
214 headerp = rdmab_to_msg(req->rl_rdmabuf);
215 headerp->rm_xid = rqst->rq_xid;
216 headerp->rm_vers = rpcrdma_version;
218 cpu_to_be32(r_xprt->rx_buf.rb_bc_srv_max_requests);
219 headerp->rm_type = rdma_msg;
220 headerp->rm_body.rm_chunks[0] = xdr_zero;
221 headerp->rm_body.rm_chunks[1] = xdr_zero;
222 headerp->rm_body.rm_chunks[2] = xdr_zero;
224 rpclen = rqst->rq_svec[0].iov_len;
226 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
227 pr_info("RPC: %s: rpclen %zd headerp 0x%p lkey 0x%x\n",
228 __func__, rpclen, headerp, rdmab_lkey(req->rl_rdmabuf));
229 pr_info("RPC: %s: RPC/RDMA: %*ph\n",
230 __func__, (int)RPCRDMA_HDRLEN_MIN, headerp);
231 pr_info("RPC: %s: RPC: %*ph\n",
232 __func__, (int)rpclen, rqst->rq_svec[0].iov_base);
235 req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
236 req->rl_send_iov[0].length = RPCRDMA_HDRLEN_MIN;
237 req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf);
239 req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf);
240 req->rl_send_iov[1].length = rpclen;
241 req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf);
248 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
249 * @xprt: transport associated with these backchannel resources
250 * @reqs: number of incoming requests to destroy; ignored
252 void xprt_rdma_bc_destroy(struct rpc_xprt *xprt, unsigned int reqs)
254 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
255 struct rpc_rqst *rqst, *tmp;
257 spin_lock_bh(&xprt->bc_pa_lock);
258 list_for_each_entry_safe(rqst, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
259 list_del(&rqst->rq_bc_pa_list);
260 spin_unlock_bh(&xprt->bc_pa_lock);
262 rpcrdma_bc_free_rqst(r_xprt, rqst);
264 spin_lock_bh(&xprt->bc_pa_lock);
266 spin_unlock_bh(&xprt->bc_pa_lock);
270 * xprt_rdma_bc_free_rqst - Release a backchannel rqst
271 * @rqst: request to release
273 void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
275 struct rpc_xprt *xprt = rqst->rq_xprt;
277 dprintk("RPC: %s: freeing rqst %p (req %p)\n",
278 __func__, rqst, rpcr_to_rdmar(rqst));
280 smp_mb__before_atomic();
281 WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state));
282 clear_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
283 smp_mb__after_atomic();
285 spin_lock_bh(&xprt->bc_pa_lock);
286 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
287 spin_unlock_bh(&xprt->bc_pa_lock);
291 * rpcrdma_bc_receive_call - Handle a backward direction call
292 * @xprt: transport receiving the call
293 * @rep: receive buffer containing the call
295 * Called in the RPC reply handler, which runs in a tasklet.
298 * Operational assumptions:
299 * o Backchannel credits are ignored, just as the NFS server
300 * forechannel currently does
301 * o The ULP manages a replay cache (eg, NFSv4.1 sessions).
302 * No replay detection is done at the transport level
304 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
305 struct rpcrdma_rep *rep)
307 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
308 struct rpcrdma_msg *headerp;
309 struct svc_serv *bc_serv;
310 struct rpcrdma_req *req;
311 struct rpc_rqst *rqst;
316 headerp = rdmab_to_msg(rep->rr_rdmabuf);
317 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
318 pr_info("RPC: %s: callback XID %08x, length=%u\n",
319 __func__, be32_to_cpu(headerp->rm_xid), rep->rr_len);
320 pr_info("RPC: %s: %*ph\n", __func__, rep->rr_len, headerp);
324 * Need at least enough bytes for RPC/RDMA header, as code
325 * here references the header fields by array offset. Also,
326 * backward calls are always inline, so ensure there
327 * are some bytes beyond the RPC/RDMA header.
329 if (rep->rr_len < RPCRDMA_HDRLEN_MIN + 24)
331 p = (__be32 *)((unsigned char *)headerp + RPCRDMA_HDRLEN_MIN);
332 size = rep->rr_len - RPCRDMA_HDRLEN_MIN;
334 /* Grab a free bc rqst */
335 spin_lock(&xprt->bc_pa_lock);
336 if (list_empty(&xprt->bc_pa_list)) {
337 spin_unlock(&xprt->bc_pa_lock);
340 rqst = list_first_entry(&xprt->bc_pa_list,
341 struct rpc_rqst, rq_bc_pa_list);
342 list_del(&rqst->rq_bc_pa_list);
343 spin_unlock(&xprt->bc_pa_lock);
344 dprintk("RPC: %s: using rqst %p\n", __func__, rqst);
347 rqst->rq_reply_bytes_recvd = 0;
348 rqst->rq_bytes_sent = 0;
349 rqst->rq_xid = headerp->rm_xid;
351 rqst->rq_private_buf.len = size;
352 set_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
354 buf = &rqst->rq_rcv_buf;
355 memset(buf, 0, sizeof(*buf));
356 buf->head[0].iov_base = p;
357 buf->head[0].iov_len = size;
360 /* The receive buffer has to be hooked to the rpcrdma_req
361 * so that it can be reposted after the server is done
362 * parsing it but just before sending the backward
365 req = rpcr_to_rdmar(rqst);
366 dprintk("RPC: %s: attaching rep %p to req %p\n",
370 /* Defeat the retransmit detection logic in send_request */
371 req->rl_connect_cookie = 0;
373 /* Queue rqst for ULP's callback service */
374 bc_serv = xprt->bc_serv;
375 spin_lock(&bc_serv->sv_cb_lock);
376 list_add(&rqst->rq_bc_list, &bc_serv->sv_cb_list);
377 spin_unlock(&bc_serv->sv_cb_lock);
379 wake_up(&bc_serv->sv_cb_waitq);
381 r_xprt->rx_stats.bcall_count++;
385 pr_warn("RPC/RDMA backchannel overflow\n");
386 xprt_disconnect_done(xprt);
387 /* This receive buffer gets reposted automatically
388 * when the connection is re-established.
393 pr_warn("RPC/RDMA short backward direction call\n");
395 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
396 xprt_disconnect_done(xprt);
398 pr_warn("RPC: %s: reposting rep %p\n",