2 * Copyright (c) 2015 Oracle. All rights reserved.
4 * Support for backward direction RPCs on RPC/RDMA (server-side).
7 #include <linux/sunrpc/svc_rdma.h>
10 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
12 #undef SVCRDMA_BACKCHANNEL_DEBUG
14 int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, struct rpcrdma_msg *rmsgp,
15 struct xdr_buf *rcvbuf)
17 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
18 struct kvec *dst, *src = &rcvbuf->head[0];
27 p = (__be32 *)src->iov_base;
31 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
32 pr_info("%s: xid=%08x, length=%zu\n",
33 __func__, be32_to_cpu(xid), len);
34 pr_info("%s: RPC/RDMA: %*ph\n",
35 __func__, (int)RPCRDMA_HDRLEN_MIN, rmsgp);
36 pr_info("%s: RPC: %*ph\n",
37 __func__, (int)len, p);
41 if (src->iov_len < 24)
44 spin_lock_bh(&xprt->transport_lock);
45 req = xprt_lookup_rqst(xprt, xid);
49 dst = &req->rq_private_buf.head[0];
50 memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf));
51 if (dst->iov_len < len)
53 memcpy(dst->iov_base, p, len);
55 credits = be32_to_cpu(rmsgp->rm_credit);
57 credits = 1; /* don't deadlock */
58 else if (credits > r_xprt->rx_buf.rb_bc_max_requests)
59 credits = r_xprt->rx_buf.rb_bc_max_requests;
62 xprt->cwnd = credits << RPC_CWNDSHIFT;
63 if (xprt->cwnd > cwnd)
64 xprt_release_rqst_cong(req->rq_task);
67 xprt_complete_rqst(req->rq_task, rcvbuf->len);
71 spin_unlock_bh(&xprt->transport_lock);
76 dprintk("svcrdma: short bc reply: xprt=%p, len=%zu\n",
81 dprintk("svcrdma: unrecognized bc reply: xprt=%p, xid=%08x\n",
82 xprt, be32_to_cpu(xid));
87 /* Send a backwards direction RPC call.
89 * Caller holds the connection's mutex and has already marshaled
90 * the RPC/RDMA request.
92 * This is similar to svc_rdma_reply, but takes an rpc_rqst
93 * instead, does not support chunks, and avoids blocking memory
96 * XXX: There is still an opportunity to block in svc_rdma_send()
97 * if there are no SQ entries to post the Send. This may occur if
98 * the adapter has a small maximum SQ depth.
100 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
101 struct rpc_rqst *rqst)
103 struct xdr_buf *sndbuf = &rqst->rq_snd_buf;
104 struct svc_rdma_op_ctxt *ctxt;
105 struct svc_rdma_req_map *vec;
106 struct ib_send_wr send_wr;
109 vec = svc_rdma_get_req_map(rdma);
110 ret = svc_rdma_map_xdr(rdma, sndbuf, vec, false);
114 ret = svc_rdma_repost_recv(rdma, GFP_NOIO);
118 ctxt = svc_rdma_get_context(rdma);
119 ctxt->pages[0] = virt_to_page(rqst->rq_buffer);
122 ctxt->direction = DMA_TO_DEVICE;
123 ctxt->sge[0].lkey = rdma->sc_pd->local_dma_lkey;
124 ctxt->sge[0].length = sndbuf->len;
126 ib_dma_map_page(rdma->sc_cm_id->device, ctxt->pages[0], 0,
127 sndbuf->len, DMA_TO_DEVICE);
128 if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr)) {
132 svc_rdma_count_mappings(rdma, ctxt);
134 memset(&send_wr, 0, sizeof(send_wr));
135 ctxt->cqe.done = svc_rdma_wc_send;
136 send_wr.wr_cqe = &ctxt->cqe;
137 send_wr.sg_list = ctxt->sge;
139 send_wr.opcode = IB_WR_SEND;
140 send_wr.send_flags = IB_SEND_SIGNALED;
142 ret = svc_rdma_send(rdma, &send_wr);
149 svc_rdma_put_req_map(rdma, vec);
150 dprintk("svcrdma: %s returns %d\n", __func__, ret);
154 svc_rdma_unmap_dma(ctxt);
155 svc_rdma_put_context(ctxt, 1);
159 /* Server-side transport endpoint wants a whole page for its send
160 * buffer. The client RPC code constructs the RPC header in this
161 * buffer before it invokes ->send_request.
164 xprt_rdma_bc_allocate(struct rpc_task *task)
166 struct rpc_rqst *rqst = task->tk_rqstp;
167 struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
168 size_t size = rqst->rq_callsize;
169 struct svcxprt_rdma *rdma;
172 rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
174 if (size > PAGE_SIZE) {
175 WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n",
180 page = alloc_page(RPCRDMA_DEF_GFP);
184 rqst->rq_buffer = page_address(page);
189 xprt_rdma_bc_free(struct rpc_task *task)
191 /* No-op: ctxt and page have already been freed. */
195 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst)
197 struct rpc_xprt *xprt = rqst->rq_xprt;
198 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
199 struct rpcrdma_msg *headerp = (struct rpcrdma_msg *)rqst->rq_buffer;
202 /* Space in the send buffer for an RPC/RDMA header is reserved
203 * via xprt->tsh_size.
205 headerp->rm_xid = rqst->rq_xid;
206 headerp->rm_vers = rpcrdma_version;
207 headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests);
208 headerp->rm_type = rdma_msg;
209 headerp->rm_body.rm_chunks[0] = xdr_zero;
210 headerp->rm_body.rm_chunks[1] = xdr_zero;
211 headerp->rm_body.rm_chunks[2] = xdr_zero;
213 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
214 pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer);
217 rc = svc_rdma_bc_sendto(rdma, rqst);
219 goto drop_connection;
223 dprintk("svcrdma: failed to send bc call\n");
224 xprt_disconnect_done(xprt);
228 /* Send an RPC call on the passive end of a transport
232 xprt_rdma_bc_send_request(struct rpc_task *task)
234 struct rpc_rqst *rqst = task->tk_rqstp;
235 struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
236 struct svcxprt_rdma *rdma;
239 dprintk("svcrdma: sending bc call with xid: %08x\n",
240 be32_to_cpu(rqst->rq_xid));
242 if (!mutex_trylock(&sxprt->xpt_mutex)) {
243 rpc_sleep_on(&sxprt->xpt_bc_pending, task, NULL);
244 if (!mutex_trylock(&sxprt->xpt_mutex))
246 rpc_wake_up_queued_task(&sxprt->xpt_bc_pending, task);
250 rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
251 if (!test_bit(XPT_DEAD, &sxprt->xpt_flags))
252 ret = rpcrdma_bc_send_request(rdma, rqst);
254 mutex_unlock(&sxprt->xpt_mutex);
262 xprt_rdma_bc_close(struct rpc_xprt *xprt)
264 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
268 xprt_rdma_bc_put(struct rpc_xprt *xprt)
270 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
273 module_put(THIS_MODULE);
276 static struct rpc_xprt_ops xprt_rdma_bc_procs = {
277 .reserve_xprt = xprt_reserve_xprt_cong,
278 .release_xprt = xprt_release_xprt_cong,
279 .alloc_slot = xprt_alloc_slot,
280 .release_request = xprt_release_rqst_cong,
281 .buf_alloc = xprt_rdma_bc_allocate,
282 .buf_free = xprt_rdma_bc_free,
283 .send_request = xprt_rdma_bc_send_request,
284 .set_retrans_timeout = xprt_set_retrans_timeout_def,
285 .close = xprt_rdma_bc_close,
286 .destroy = xprt_rdma_bc_put,
287 .print_stats = xprt_rdma_print_stats
290 static const struct rpc_timeout xprt_rdma_bc_timeout = {
291 .to_initval = 60 * HZ,
292 .to_maxval = 60 * HZ,
295 /* It shouldn't matter if the number of backchannel session slots
296 * doesn't match the number of RPC/RDMA credits. That just means
297 * one or the other will have extra slots that aren't used.
299 static struct rpc_xprt *
300 xprt_setup_rdma_bc(struct xprt_create *args)
302 struct rpc_xprt *xprt;
303 struct rpcrdma_xprt *new_xprt;
305 if (args->addrlen > sizeof(xprt->addr)) {
306 dprintk("RPC: %s: address too large\n", __func__);
307 return ERR_PTR(-EBADF);
310 xprt = xprt_alloc(args->net, sizeof(*new_xprt),
311 RPCRDMA_MAX_BC_REQUESTS,
312 RPCRDMA_MAX_BC_REQUESTS);
314 dprintk("RPC: %s: couldn't allocate rpc_xprt\n",
316 return ERR_PTR(-ENOMEM);
319 xprt->timeout = &xprt_rdma_bc_timeout;
320 xprt_set_bound(xprt);
321 xprt_set_connected(xprt);
322 xprt->bind_timeout = RPCRDMA_BIND_TO;
323 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
324 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
326 xprt->prot = XPRT_TRANSPORT_BC_RDMA;
327 xprt->tsh_size = RPCRDMA_HDRLEN_MIN / sizeof(__be32);
328 xprt->ops = &xprt_rdma_bc_procs;
330 memcpy(&xprt->addr, args->dstaddr, args->addrlen);
331 xprt->addrlen = args->addrlen;
332 xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr);
335 xprt->max_payload = xprt_rdma_max_inline_read;
337 new_xprt = rpcx_to_rdmax(xprt);
338 new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs;
341 args->bc_xprt->xpt_bc_xprt = xprt;
342 xprt->bc_xprt = args->bc_xprt;
344 if (!try_module_get(THIS_MODULE))
347 /* Final put for backchannel xprt is in __svc_rdma_free */
352 xprt_rdma_free_addresses(xprt);
353 args->bc_xprt->xpt_bc_xprt = NULL;
356 return ERR_PTR(-EINVAL);
359 struct xprt_class xprt_rdma_bc = {
360 .list = LIST_HEAD_INIT(xprt_rdma_bc.list),
361 .name = "rdma backchannel",
362 .owner = THIS_MODULE,
363 .ident = XPRT_TRANSPORT_BC_RDMA,
364 .setup = xprt_setup_rdma_bc,