7e188bc793a200eb2be359232ff93c7ea35bd223
[muen/linux.git] / net / sunrpc / xprtrdma / svc_rdma_backchannel.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2015-2018 Oracle.  All rights reserved.
4  *
5  * Support for backward direction RPCs on RPC/RDMA (server-side).
6  */
7
8 #include <linux/sunrpc/svc_rdma.h>
9
10 #include "xprt_rdma.h"
11 #include <trace/events/rpcrdma.h>
12
13 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
14
15 #undef SVCRDMA_BACKCHANNEL_DEBUG
16
17 /**
18  * svc_rdma_handle_bc_reply - Process incoming backchannel reply
19  * @xprt: controlling backchannel transport
20  * @rdma_resp: pointer to incoming transport header
21  * @rcvbuf: XDR buffer into which to decode the reply
22  *
23  * Returns:
24  *      %0 if @rcvbuf is filled in, xprt_complete_rqst called,
25  *      %-EAGAIN if server should call ->recvfrom again.
26  */
27 int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, __be32 *rdma_resp,
28                              struct xdr_buf *rcvbuf)
29 {
30         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
31         struct kvec *dst, *src = &rcvbuf->head[0];
32         struct rpc_rqst *req;
33         u32 credits;
34         size_t len;
35         __be32 xid;
36         __be32 *p;
37         int ret;
38
39         p = (__be32 *)src->iov_base;
40         len = src->iov_len;
41         xid = *rdma_resp;
42
43 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
44         pr_info("%s: xid=%08x, length=%zu\n",
45                 __func__, be32_to_cpu(xid), len);
46         pr_info("%s: RPC/RDMA: %*ph\n",
47                 __func__, (int)RPCRDMA_HDRLEN_MIN, rdma_resp);
48         pr_info("%s:      RPC: %*ph\n",
49                 __func__, (int)len, p);
50 #endif
51
52         ret = -EAGAIN;
53         if (src->iov_len < 24)
54                 goto out_shortreply;
55
56         spin_lock(&xprt->recv_lock);
57         req = xprt_lookup_rqst(xprt, xid);
58         if (!req)
59                 goto out_notfound;
60
61         dst = &req->rq_private_buf.head[0];
62         memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf));
63         if (dst->iov_len < len)
64                 goto out_unlock;
65         memcpy(dst->iov_base, p, len);
66         xprt_pin_rqst(req);
67         spin_unlock(&xprt->recv_lock);
68
69         credits = be32_to_cpup(rdma_resp + 2);
70         if (credits == 0)
71                 credits = 1;    /* don't deadlock */
72         else if (credits > r_xprt->rx_buf.rb_bc_max_requests)
73                 credits = r_xprt->rx_buf.rb_bc_max_requests;
74
75         spin_lock_bh(&xprt->transport_lock);
76         xprt->cwnd = credits << RPC_CWNDSHIFT;
77         spin_unlock_bh(&xprt->transport_lock);
78
79         spin_lock(&xprt->recv_lock);
80         ret = 0;
81         xprt_complete_rqst(req->rq_task, rcvbuf->len);
82         xprt_unpin_rqst(req);
83         rcvbuf->len = 0;
84
85 out_unlock:
86         spin_unlock(&xprt->recv_lock);
87 out:
88         return ret;
89
90 out_shortreply:
91         dprintk("svcrdma: short bc reply: xprt=%p, len=%zu\n",
92                 xprt, src->iov_len);
93         goto out;
94
95 out_notfound:
96         dprintk("svcrdma: unrecognized bc reply: xprt=%p, xid=%08x\n",
97                 xprt, be32_to_cpu(xid));
98         goto out_unlock;
99 }
100
101 /* Send a backwards direction RPC call.
102  *
103  * Caller holds the connection's mutex and has already marshaled
104  * the RPC/RDMA request.
105  *
106  * This is similar to svc_rdma_send_reply_msg, but takes a struct
107  * rpc_rqst instead, does not support chunks, and avoids blocking
108  * memory allocation.
109  *
110  * XXX: There is still an opportunity to block in svc_rdma_send()
111  * if there are no SQ entries to post the Send. This may occur if
112  * the adapter has a small maximum SQ depth.
113  */
114 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
115                               struct rpc_rqst *rqst,
116                               struct svc_rdma_send_ctxt *ctxt)
117 {
118         int ret;
119
120         ret = svc_rdma_map_reply_msg(rdma, ctxt, &rqst->rq_snd_buf, NULL);
121         if (ret < 0)
122                 return -EIO;
123
124         /* Bump page refcnt so Send completion doesn't release
125          * the rq_buffer before all retransmits are complete.
126          */
127         get_page(virt_to_page(rqst->rq_buffer));
128         ctxt->sc_send_wr.opcode = IB_WR_SEND;
129         return svc_rdma_send(rdma, &ctxt->sc_send_wr);
130 }
131
132 /* Server-side transport endpoint wants a whole page for its send
133  * buffer. The client RPC code constructs the RPC header in this
134  * buffer before it invokes ->send_request.
135  */
136 static int
137 xprt_rdma_bc_allocate(struct rpc_task *task)
138 {
139         struct rpc_rqst *rqst = task->tk_rqstp;
140         size_t size = rqst->rq_callsize;
141         struct page *page;
142
143         if (size > PAGE_SIZE) {
144                 WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n",
145                           size);
146                 return -EINVAL;
147         }
148
149         page = alloc_page(RPCRDMA_DEF_GFP);
150         if (!page)
151                 return -ENOMEM;
152         rqst->rq_buffer = page_address(page);
153
154         rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, RPCRDMA_DEF_GFP);
155         if (!rqst->rq_rbuffer) {
156                 put_page(page);
157                 return -ENOMEM;
158         }
159         return 0;
160 }
161
162 static void
163 xprt_rdma_bc_free(struct rpc_task *task)
164 {
165         struct rpc_rqst *rqst = task->tk_rqstp;
166
167         put_page(virt_to_page(rqst->rq_buffer));
168         kfree(rqst->rq_rbuffer);
169 }
170
171 static int
172 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst)
173 {
174         struct rpc_xprt *xprt = rqst->rq_xprt;
175         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
176         struct svc_rdma_send_ctxt *ctxt;
177         __be32 *p;
178         int rc;
179
180         ctxt = svc_rdma_send_ctxt_get(rdma);
181         if (!ctxt)
182                 goto drop_connection;
183
184         p = ctxt->sc_xprt_buf;
185         *p++ = rqst->rq_xid;
186         *p++ = rpcrdma_version;
187         *p++ = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests);
188         *p++ = rdma_msg;
189         *p++ = xdr_zero;
190         *p++ = xdr_zero;
191         *p   = xdr_zero;
192         svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_MIN);
193
194 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
195         pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer);
196 #endif
197
198         rc = svc_rdma_bc_sendto(rdma, rqst, ctxt);
199         if (rc) {
200                 svc_rdma_send_ctxt_put(rdma, ctxt);
201                 goto drop_connection;
202         }
203         return rc;
204
205 drop_connection:
206         dprintk("svcrdma: failed to send bc call\n");
207         xprt_disconnect_done(xprt);
208         return -ENOTCONN;
209 }
210
211 /* Send an RPC call on the passive end of a transport
212  * connection.
213  */
214 static int
215 xprt_rdma_bc_send_request(struct rpc_task *task)
216 {
217         struct rpc_rqst *rqst = task->tk_rqstp;
218         struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
219         struct svcxprt_rdma *rdma;
220         int ret;
221
222         dprintk("svcrdma: sending bc call with xid: %08x\n",
223                 be32_to_cpu(rqst->rq_xid));
224
225         if (!mutex_trylock(&sxprt->xpt_mutex)) {
226                 rpc_sleep_on(&sxprt->xpt_bc_pending, task, NULL);
227                 if (!mutex_trylock(&sxprt->xpt_mutex))
228                         return -EAGAIN;
229                 rpc_wake_up_queued_task(&sxprt->xpt_bc_pending, task);
230         }
231
232         ret = -ENOTCONN;
233         rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
234         if (!test_bit(XPT_DEAD, &sxprt->xpt_flags))
235                 ret = rpcrdma_bc_send_request(rdma, rqst);
236
237         mutex_unlock(&sxprt->xpt_mutex);
238
239         if (ret < 0)
240                 return ret;
241         return 0;
242 }
243
244 static void
245 xprt_rdma_bc_close(struct rpc_xprt *xprt)
246 {
247         dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
248 }
249
250 static void
251 xprt_rdma_bc_put(struct rpc_xprt *xprt)
252 {
253         dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
254
255         xprt_free(xprt);
256 }
257
258 static const struct rpc_xprt_ops xprt_rdma_bc_procs = {
259         .reserve_xprt           = xprt_reserve_xprt_cong,
260         .release_xprt           = xprt_release_xprt_cong,
261         .alloc_slot             = xprt_alloc_slot,
262         .free_slot              = xprt_free_slot,
263         .release_request        = xprt_release_rqst_cong,
264         .buf_alloc              = xprt_rdma_bc_allocate,
265         .buf_free               = xprt_rdma_bc_free,
266         .send_request           = xprt_rdma_bc_send_request,
267         .set_retrans_timeout    = xprt_set_retrans_timeout_def,
268         .close                  = xprt_rdma_bc_close,
269         .destroy                = xprt_rdma_bc_put,
270         .print_stats            = xprt_rdma_print_stats
271 };
272
273 static const struct rpc_timeout xprt_rdma_bc_timeout = {
274         .to_initval = 60 * HZ,
275         .to_maxval = 60 * HZ,
276 };
277
278 /* It shouldn't matter if the number of backchannel session slots
279  * doesn't match the number of RPC/RDMA credits. That just means
280  * one or the other will have extra slots that aren't used.
281  */
282 static struct rpc_xprt *
283 xprt_setup_rdma_bc(struct xprt_create *args)
284 {
285         struct rpc_xprt *xprt;
286         struct rpcrdma_xprt *new_xprt;
287
288         if (args->addrlen > sizeof(xprt->addr)) {
289                 dprintk("RPC:       %s: address too large\n", __func__);
290                 return ERR_PTR(-EBADF);
291         }
292
293         xprt = xprt_alloc(args->net, sizeof(*new_xprt),
294                           RPCRDMA_MAX_BC_REQUESTS,
295                           RPCRDMA_MAX_BC_REQUESTS);
296         if (!xprt) {
297                 dprintk("RPC:       %s: couldn't allocate rpc_xprt\n",
298                         __func__);
299                 return ERR_PTR(-ENOMEM);
300         }
301
302         xprt->timeout = &xprt_rdma_bc_timeout;
303         xprt_set_bound(xprt);
304         xprt_set_connected(xprt);
305         xprt->bind_timeout = RPCRDMA_BIND_TO;
306         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
307         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
308
309         xprt->prot = XPRT_TRANSPORT_BC_RDMA;
310         xprt->tsh_size = 0;
311         xprt->ops = &xprt_rdma_bc_procs;
312
313         memcpy(&xprt->addr, args->dstaddr, args->addrlen);
314         xprt->addrlen = args->addrlen;
315         xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr);
316         xprt->resvport = 0;
317
318         xprt->max_payload = xprt_rdma_max_inline_read;
319
320         new_xprt = rpcx_to_rdmax(xprt);
321         new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs;
322
323         xprt_get(xprt);
324         args->bc_xprt->xpt_bc_xprt = xprt;
325         xprt->bc_xprt = args->bc_xprt;
326
327         /* Final put for backchannel xprt is in __svc_rdma_free */
328         xprt_get(xprt);
329         return xprt;
330 }
331
332 struct xprt_class xprt_rdma_bc = {
333         .list                   = LIST_HEAD_INIT(xprt_rdma_bc.list),
334         .name                   = "rdma backchannel",
335         .owner                  = THIS_MODULE,
336         .ident                  = XPRT_TRANSPORT_BC_RDMA,
337         .setup                  = xprt_setup_rdma_bc,
338 };