1 /*
2 * Copyright (c) 2015 Oracle. All rights reserved.
3 *
4 * Support for backward direction RPCs on RPC/RDMA (server-side).
5 */
6
7 #include <linux/sunrpc/svc_rdma.h>
8 #include "xprt_rdma.h"
9
10 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
11
12 #undef SVCRDMA_BACKCHANNEL_DEBUG
13
svc_rdma_handle_bc_reply(struct rpc_xprt * xprt,struct rpcrdma_msg * rmsgp,struct xdr_buf * rcvbuf)14 int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, struct rpcrdma_msg *rmsgp,
15 struct xdr_buf *rcvbuf)
16 {
17 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
18 struct kvec *dst, *src = &rcvbuf->head[0];
19 struct rpc_rqst *req;
20 unsigned long cwnd;
21 u32 credits;
22 size_t len;
23 __be32 xid;
24 __be32 *p;
25 int ret;
26
27 p = (__be32 *)src->iov_base;
28 len = src->iov_len;
29 xid = rmsgp->rm_xid;
30
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);
38 #endif
39
40 ret = -EAGAIN;
41 if (src->iov_len < 24)
42 goto out_shortreply;
43
44 spin_lock_bh(&xprt->transport_lock);
45 req = xprt_lookup_rqst(xprt, xid);
46 if (!req)
47 goto out_notfound;
48
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)
52 goto out_unlock;
53 memcpy(dst->iov_base, p, len);
54
55 credits = be32_to_cpu(rmsgp->rm_credit);
56 if (credits == 0)
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;
60
61 cwnd = xprt->cwnd;
62 xprt->cwnd = credits << RPC_CWNDSHIFT;
63 if (xprt->cwnd > cwnd)
64 xprt_release_rqst_cong(req->rq_task);
65
66 ret = 0;
67 xprt_complete_rqst(req->rq_task, rcvbuf->len);
68 rcvbuf->len = 0;
69
70 out_unlock:
71 spin_unlock_bh(&xprt->transport_lock);
72 out:
73 return ret;
74
75 out_shortreply:
76 dprintk("svcrdma: short bc reply: xprt=%p, len=%zu\n",
77 xprt, src->iov_len);
78 goto out;
79
80 out_notfound:
81 dprintk("svcrdma: unrecognized bc reply: xprt=%p, xid=%08x\n",
82 xprt, be32_to_cpu(xid));
83
84 goto out_unlock;
85 }
86
87 /* Send a backwards direction RPC call.
88 *
89 * Caller holds the connection's mutex and has already marshaled
90 * the RPC/RDMA request.
91 *
92 * This is similar to svc_rdma_reply, but takes an rpc_rqst
93 * instead, does not support chunks, and avoids blocking memory
94 * allocation.
95 *
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.
99 */
svc_rdma_bc_sendto(struct svcxprt_rdma * rdma,struct rpc_rqst * rqst)100 static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
101 struct rpc_rqst *rqst)
102 {
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;
107 int ret;
108
109 vec = svc_rdma_get_req_map(rdma);
110 ret = svc_rdma_map_xdr(rdma, sndbuf, vec, false);
111 if (ret)
112 goto out_err;
113
114 ret = svc_rdma_repost_recv(rdma, GFP_NOIO);
115 if (ret)
116 goto out_err;
117
118 ctxt = svc_rdma_get_context(rdma);
119 ctxt->pages[0] = virt_to_page(rqst->rq_buffer);
120 ctxt->count = 1;
121
122 ctxt->direction = DMA_TO_DEVICE;
123 ctxt->sge[0].lkey = rdma->sc_pd->local_dma_lkey;
124 ctxt->sge[0].length = sndbuf->len;
125 ctxt->sge[0].addr =
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)) {
129 ret = -EIO;
130 goto out_unmap;
131 }
132 svc_rdma_count_mappings(rdma, ctxt);
133
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;
138 send_wr.num_sge = 1;
139 send_wr.opcode = IB_WR_SEND;
140 send_wr.send_flags = IB_SEND_SIGNALED;
141
142 ret = svc_rdma_send(rdma, &send_wr);
143 if (ret) {
144 ret = -EIO;
145 goto out_unmap;
146 }
147
148 out_err:
149 svc_rdma_put_req_map(rdma, vec);
150 dprintk("svcrdma: %s returns %d\n", __func__, ret);
151 return ret;
152
153 out_unmap:
154 svc_rdma_unmap_dma(ctxt);
155 svc_rdma_put_context(ctxt, 1);
156 goto out_err;
157 }
158
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.
162 */
163 static int
xprt_rdma_bc_allocate(struct rpc_task * task)164 xprt_rdma_bc_allocate(struct rpc_task *task)
165 {
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;
170 struct page *page;
171
172 rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
173
174 if (size > PAGE_SIZE) {
175 WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n",
176 size);
177 return -EINVAL;
178 }
179
180 /* svc_rdma_sendto releases this page */
181 page = alloc_page(RPCRDMA_DEF_GFP);
182 if (!page)
183 return -ENOMEM;
184 rqst->rq_buffer = page_address(page);
185
186 rqst->rq_rbuffer = kmalloc(rqst->rq_rcvsize, RPCRDMA_DEF_GFP);
187 if (!rqst->rq_rbuffer) {
188 put_page(page);
189 return -ENOMEM;
190 }
191 return 0;
192 }
193
194 static void
xprt_rdma_bc_free(struct rpc_task * task)195 xprt_rdma_bc_free(struct rpc_task *task)
196 {
197 struct rpc_rqst *rqst = task->tk_rqstp;
198
199 kfree(rqst->rq_rbuffer);
200 }
201
202 static int
rpcrdma_bc_send_request(struct svcxprt_rdma * rdma,struct rpc_rqst * rqst)203 rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst)
204 {
205 struct rpc_xprt *xprt = rqst->rq_xprt;
206 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
207 struct rpcrdma_msg *headerp = (struct rpcrdma_msg *)rqst->rq_buffer;
208 int rc;
209
210 /* Space in the send buffer for an RPC/RDMA header is reserved
211 * via xprt->tsh_size.
212 */
213 headerp->rm_xid = rqst->rq_xid;
214 headerp->rm_vers = rpcrdma_version;
215 headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests);
216 headerp->rm_type = rdma_msg;
217 headerp->rm_body.rm_chunks[0] = xdr_zero;
218 headerp->rm_body.rm_chunks[1] = xdr_zero;
219 headerp->rm_body.rm_chunks[2] = xdr_zero;
220
221 #ifdef SVCRDMA_BACKCHANNEL_DEBUG
222 pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer);
223 #endif
224
225 rc = svc_rdma_bc_sendto(rdma, rqst);
226 if (rc)
227 goto drop_connection;
228 return rc;
229
230 drop_connection:
231 dprintk("svcrdma: failed to send bc call\n");
232 xprt_disconnect_done(xprt);
233 return -ENOTCONN;
234 }
235
236 /* Send an RPC call on the passive end of a transport
237 * connection.
238 */
239 static int
xprt_rdma_bc_send_request(struct rpc_task * task)240 xprt_rdma_bc_send_request(struct rpc_task *task)
241 {
242 struct rpc_rqst *rqst = task->tk_rqstp;
243 struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
244 struct svcxprt_rdma *rdma;
245 int ret;
246
247 dprintk("svcrdma: sending bc call with xid: %08x\n",
248 be32_to_cpu(rqst->rq_xid));
249
250 if (!mutex_trylock(&sxprt->xpt_mutex)) {
251 rpc_sleep_on(&sxprt->xpt_bc_pending, task, NULL);
252 if (!mutex_trylock(&sxprt->xpt_mutex))
253 return -EAGAIN;
254 rpc_wake_up_queued_task(&sxprt->xpt_bc_pending, task);
255 }
256
257 ret = -ENOTCONN;
258 rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
259 if (!test_bit(XPT_DEAD, &sxprt->xpt_flags))
260 ret = rpcrdma_bc_send_request(rdma, rqst);
261
262 mutex_unlock(&sxprt->xpt_mutex);
263
264 if (ret < 0)
265 return ret;
266 return 0;
267 }
268
269 static void
xprt_rdma_bc_close(struct rpc_xprt * xprt)270 xprt_rdma_bc_close(struct rpc_xprt *xprt)
271 {
272 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
273 }
274
275 static void
xprt_rdma_bc_put(struct rpc_xprt * xprt)276 xprt_rdma_bc_put(struct rpc_xprt *xprt)
277 {
278 dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
279
280 xprt_free(xprt);
281 module_put(THIS_MODULE);
282 }
283
284 static struct rpc_xprt_ops xprt_rdma_bc_procs = {
285 .reserve_xprt = xprt_reserve_xprt_cong,
286 .release_xprt = xprt_release_xprt_cong,
287 .alloc_slot = xprt_alloc_slot,
288 .release_request = xprt_release_rqst_cong,
289 .buf_alloc = xprt_rdma_bc_allocate,
290 .buf_free = xprt_rdma_bc_free,
291 .send_request = xprt_rdma_bc_send_request,
292 .set_retrans_timeout = xprt_set_retrans_timeout_def,
293 .close = xprt_rdma_bc_close,
294 .destroy = xprt_rdma_bc_put,
295 .print_stats = xprt_rdma_print_stats
296 };
297
298 static const struct rpc_timeout xprt_rdma_bc_timeout = {
299 .to_initval = 60 * HZ,
300 .to_maxval = 60 * HZ,
301 };
302
303 /* It shouldn't matter if the number of backchannel session slots
304 * doesn't match the number of RPC/RDMA credits. That just means
305 * one or the other will have extra slots that aren't used.
306 */
307 static struct rpc_xprt *
xprt_setup_rdma_bc(struct xprt_create * args)308 xprt_setup_rdma_bc(struct xprt_create *args)
309 {
310 struct rpc_xprt *xprt;
311 struct rpcrdma_xprt *new_xprt;
312
313 if (args->addrlen > sizeof(xprt->addr)) {
314 dprintk("RPC: %s: address too large\n", __func__);
315 return ERR_PTR(-EBADF);
316 }
317
318 xprt = xprt_alloc(args->net, sizeof(*new_xprt),
319 RPCRDMA_MAX_BC_REQUESTS,
320 RPCRDMA_MAX_BC_REQUESTS);
321 if (!xprt) {
322 dprintk("RPC: %s: couldn't allocate rpc_xprt\n",
323 __func__);
324 return ERR_PTR(-ENOMEM);
325 }
326
327 xprt->timeout = &xprt_rdma_bc_timeout;
328 xprt_set_bound(xprt);
329 xprt_set_connected(xprt);
330 xprt->bind_timeout = RPCRDMA_BIND_TO;
331 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
332 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
333
334 xprt->prot = XPRT_TRANSPORT_BC_RDMA;
335 xprt->tsh_size = RPCRDMA_HDRLEN_MIN / sizeof(__be32);
336 xprt->ops = &xprt_rdma_bc_procs;
337
338 memcpy(&xprt->addr, args->dstaddr, args->addrlen);
339 xprt->addrlen = args->addrlen;
340 xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr);
341 xprt->resvport = 0;
342
343 xprt->max_payload = xprt_rdma_max_inline_read;
344
345 new_xprt = rpcx_to_rdmax(xprt);
346 new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs;
347
348 xprt_get(xprt);
349 args->bc_xprt->xpt_bc_xprt = xprt;
350 xprt->bc_xprt = args->bc_xprt;
351
352 if (!try_module_get(THIS_MODULE))
353 goto out_fail;
354
355 /* Final put for backchannel xprt is in __svc_rdma_free */
356 xprt_get(xprt);
357 return xprt;
358
359 out_fail:
360 xprt_rdma_free_addresses(xprt);
361 args->bc_xprt->xpt_bc_xprt = NULL;
362 args->bc_xprt->xpt_bc_xps = NULL;
363 xprt_put(xprt);
364 xprt_free(xprt);
365 return ERR_PTR(-EINVAL);
366 }
367
368 struct xprt_class xprt_rdma_bc = {
369 .list = LIST_HEAD_INIT(xprt_rdma_bc.list),
370 .name = "rdma backchannel",
371 .owner = THIS_MODULE,
372 .ident = XPRT_TRANSPORT_BC_RDMA,
373 .setup = xprt_setup_rdma_bc,
374 };
375