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(req
);
35 static int rpcrdma_bc_setup_rqst(struct rpcrdma_xprt
*r_xprt
,
36 struct rpc_rqst
*rqst
)
38 struct rpcrdma_regbuf
*rb
;
39 struct rpcrdma_req
*req
;
42 req
= rpcrdma_create_req(r_xprt
);
45 req
->rl_backchannel
= true;
47 rb
= rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE
,
48 DMA_TO_DEVICE
, GFP_KERNEL
);
53 size
= r_xprt
->rx_data
.inline_rsize
;
54 rb
= rpcrdma_alloc_regbuf(size
, DMA_TO_DEVICE
, GFP_KERNEL
);
58 xdr_buf_init(&rqst
->rq_snd_buf
, rb
->rg_base
,
59 min_t(size_t, size
, PAGE_SIZE
));
60 rpcrdma_set_xprtdata(rqst
, req
);
64 rpcrdma_bc_free_rqst(r_xprt
, rqst
);
68 /* Allocate and add receive buffers to the rpcrdma_buffer's
69 * existing list of rep's. These are released when the
70 * transport is destroyed.
72 static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt
*r_xprt
,
75 struct rpcrdma_rep
*rep
;
79 rep
= rpcrdma_create_rep(r_xprt
);
81 pr_err("RPC: %s: reply buffer alloc failed\n",
87 rpcrdma_recv_buffer_put(rep
);
94 * xprt_rdma_bc_setup - Pre-allocate resources for handling backchannel requests
95 * @xprt: transport associated with these backchannel resources
96 * @reqs: number of concurrent incoming requests to expect
98 * Returns 0 on success; otherwise a negative errno
100 int xprt_rdma_bc_setup(struct rpc_xprt
*xprt
, unsigned int reqs
)
102 struct rpcrdma_xprt
*r_xprt
= rpcx_to_rdmax(xprt
);
103 struct rpcrdma_buffer
*buffer
= &r_xprt
->rx_buf
;
104 struct rpc_rqst
*rqst
;
108 /* The backchannel reply path returns each rpc_rqst to the
109 * bc_pa_list _after_ the reply is sent. If the server is
110 * faster than the client, it can send another backward
111 * direction request before the rpc_rqst is returned to the
112 * list. The client rejects the request in this case.
114 * Twice as many rpc_rqsts are prepared to ensure there is
115 * always an rpc_rqst available as soon as a reply is sent.
117 if (reqs
> RPCRDMA_BACKWARD_WRS
>> 1)
120 for (i
= 0; i
< (reqs
<< 1); i
++) {
121 rqst
= kzalloc(sizeof(*rqst
), GFP_KERNEL
);
125 dprintk("RPC: %s: new rqst %p\n", __func__
, rqst
);
127 rqst
->rq_xprt
= &r_xprt
->rx_xprt
;
128 INIT_LIST_HEAD(&rqst
->rq_list
);
129 INIT_LIST_HEAD(&rqst
->rq_bc_list
);
131 if (rpcrdma_bc_setup_rqst(r_xprt
, rqst
))
134 spin_lock_bh(&xprt
->bc_pa_lock
);
135 list_add(&rqst
->rq_bc_pa_list
, &xprt
->bc_pa_list
);
136 spin_unlock_bh(&xprt
->bc_pa_lock
);
139 rc
= rpcrdma_bc_setup_reps(r_xprt
, reqs
);
143 rc
= rpcrdma_ep_post_extra_recv(r_xprt
, reqs
);
147 buffer
->rb_bc_srv_max_requests
= reqs
;
148 request_module("svcrdma");
153 xprt_rdma_bc_destroy(xprt
, reqs
);
156 pr_err("RPC: %s: setup backchannel transport failed\n", __func__
);
161 * xprt_rdma_bc_up - Create transport endpoint for backchannel service
162 * @serv: server endpoint
163 * @net: network namespace
165 * The "xprt" is an implied argument: it supplies the name of the
166 * backchannel transport class.
168 * Returns zero on success, negative errno on failure
170 int xprt_rdma_bc_up(struct svc_serv
*serv
, struct net
*net
)
174 ret
= svc_create_xprt(serv
, "rdma-bc", net
, PF_INET
, 0, 0);
181 * xprt_rdma_bc_maxpayload - Return maximum backchannel message size
184 * Returns maximum size, in bytes, of a backchannel message
186 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt
*xprt
)
188 struct rpcrdma_xprt
*r_xprt
= rpcx_to_rdmax(xprt
);
189 struct rpcrdma_create_data_internal
*cdata
= &r_xprt
->rx_data
;
192 maxmsg
= min_t(unsigned int, cdata
->inline_rsize
, cdata
->inline_wsize
);
193 maxmsg
= min_t(unsigned int, maxmsg
, PAGE_SIZE
);
194 return maxmsg
- RPCRDMA_HDRLEN_MIN
;
198 * rpcrdma_bc_marshal_reply - Send backwards direction reply
199 * @rqst: buffer containing RPC reply data
201 * Returns zero on success.
203 int rpcrdma_bc_marshal_reply(struct rpc_rqst
*rqst
)
205 struct rpc_xprt
*xprt
= rqst
->rq_xprt
;
206 struct rpcrdma_xprt
*r_xprt
= rpcx_to_rdmax(xprt
);
207 struct rpcrdma_req
*req
= rpcr_to_rdmar(rqst
);
208 struct rpcrdma_msg
*headerp
;
210 headerp
= rdmab_to_msg(req
->rl_rdmabuf
);
211 headerp
->rm_xid
= rqst
->rq_xid
;
212 headerp
->rm_vers
= rpcrdma_version
;
214 cpu_to_be32(r_xprt
->rx_buf
.rb_bc_srv_max_requests
);
215 headerp
->rm_type
= rdma_msg
;
216 headerp
->rm_body
.rm_chunks
[0] = xdr_zero
;
217 headerp
->rm_body
.rm_chunks
[1] = xdr_zero
;
218 headerp
->rm_body
.rm_chunks
[2] = xdr_zero
;
220 if (!rpcrdma_prepare_send_sges(&r_xprt
->rx_ia
, req
, RPCRDMA_HDRLEN_MIN
,
221 &rqst
->rq_snd_buf
, rpcrdma_noch
))
227 * xprt_rdma_bc_destroy - Release resources for handling backchannel requests
228 * @xprt: transport associated with these backchannel resources
229 * @reqs: number of incoming requests to destroy; ignored
231 void xprt_rdma_bc_destroy(struct rpc_xprt
*xprt
, unsigned int reqs
)
233 struct rpcrdma_xprt
*r_xprt
= rpcx_to_rdmax(xprt
);
234 struct rpc_rqst
*rqst
, *tmp
;
236 spin_lock_bh(&xprt
->bc_pa_lock
);
237 list_for_each_entry_safe(rqst
, tmp
, &xprt
->bc_pa_list
, rq_bc_pa_list
) {
238 list_del(&rqst
->rq_bc_pa_list
);
239 spin_unlock_bh(&xprt
->bc_pa_lock
);
241 rpcrdma_bc_free_rqst(r_xprt
, rqst
);
243 spin_lock_bh(&xprt
->bc_pa_lock
);
245 spin_unlock_bh(&xprt
->bc_pa_lock
);
249 * xprt_rdma_bc_free_rqst - Release a backchannel rqst
250 * @rqst: request to release
252 void xprt_rdma_bc_free_rqst(struct rpc_rqst
*rqst
)
254 struct rpc_xprt
*xprt
= rqst
->rq_xprt
;
256 dprintk("RPC: %s: freeing rqst %p (req %p)\n",
257 __func__
, rqst
, rpcr_to_rdmar(rqst
));
259 smp_mb__before_atomic();
260 WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE
, &rqst
->rq_bc_pa_state
));
261 clear_bit(RPC_BC_PA_IN_USE
, &rqst
->rq_bc_pa_state
);
262 smp_mb__after_atomic();
264 spin_lock_bh(&xprt
->bc_pa_lock
);
265 list_add_tail(&rqst
->rq_bc_pa_list
, &xprt
->bc_pa_list
);
266 spin_unlock_bh(&xprt
->bc_pa_lock
);
270 * rpcrdma_bc_receive_call - Handle a backward direction call
271 * @xprt: transport receiving the call
272 * @rep: receive buffer containing the call
274 * Called in the RPC reply handler, which runs in a tasklet.
277 * Operational assumptions:
278 * o Backchannel credits are ignored, just as the NFS server
279 * forechannel currently does
280 * o The ULP manages a replay cache (eg, NFSv4.1 sessions).
281 * No replay detection is done at the transport level
283 void rpcrdma_bc_receive_call(struct rpcrdma_xprt
*r_xprt
,
284 struct rpcrdma_rep
*rep
)
286 struct rpc_xprt
*xprt
= &r_xprt
->rx_xprt
;
287 struct rpcrdma_msg
*headerp
;
288 struct svc_serv
*bc_serv
;
289 struct rpcrdma_req
*req
;
290 struct rpc_rqst
*rqst
;
295 headerp
= rdmab_to_msg(rep
->rr_rdmabuf
);
296 #ifdef RPCRDMA_BACKCHANNEL_DEBUG
297 pr_info("RPC: %s: callback XID %08x, length=%u\n",
298 __func__
, be32_to_cpu(headerp
->rm_xid
), rep
->rr_len
);
299 pr_info("RPC: %s: %*ph\n", __func__
, rep
->rr_len
, headerp
);
303 * Need at least enough bytes for RPC/RDMA header, as code
304 * here references the header fields by array offset. Also,
305 * backward calls are always inline, so ensure there
306 * are some bytes beyond the RPC/RDMA header.
308 if (rep
->rr_len
< RPCRDMA_HDRLEN_MIN
+ 24)
310 p
= (__be32
*)((unsigned char *)headerp
+ RPCRDMA_HDRLEN_MIN
);
311 size
= rep
->rr_len
- RPCRDMA_HDRLEN_MIN
;
313 /* Grab a free bc rqst */
314 spin_lock(&xprt
->bc_pa_lock
);
315 if (list_empty(&xprt
->bc_pa_list
)) {
316 spin_unlock(&xprt
->bc_pa_lock
);
319 rqst
= list_first_entry(&xprt
->bc_pa_list
,
320 struct rpc_rqst
, rq_bc_pa_list
);
321 list_del(&rqst
->rq_bc_pa_list
);
322 spin_unlock(&xprt
->bc_pa_lock
);
323 dprintk("RPC: %s: using rqst %p\n", __func__
, rqst
);
326 rqst
->rq_reply_bytes_recvd
= 0;
327 rqst
->rq_bytes_sent
= 0;
328 rqst
->rq_xid
= headerp
->rm_xid
;
330 rqst
->rq_private_buf
.len
= size
;
331 set_bit(RPC_BC_PA_IN_USE
, &rqst
->rq_bc_pa_state
);
333 buf
= &rqst
->rq_rcv_buf
;
334 memset(buf
, 0, sizeof(*buf
));
335 buf
->head
[0].iov_base
= p
;
336 buf
->head
[0].iov_len
= size
;
339 /* The receive buffer has to be hooked to the rpcrdma_req
340 * so that it can be reposted after the server is done
341 * parsing it but just before sending the backward
344 req
= rpcr_to_rdmar(rqst
);
345 dprintk("RPC: %s: attaching rep %p to req %p\n",
349 /* Defeat the retransmit detection logic in send_request */
350 req
->rl_connect_cookie
= 0;
352 /* Queue rqst for ULP's callback service */
353 bc_serv
= xprt
->bc_serv
;
354 spin_lock(&bc_serv
->sv_cb_lock
);
355 list_add(&rqst
->rq_bc_list
, &bc_serv
->sv_cb_list
);
356 spin_unlock(&bc_serv
->sv_cb_lock
);
358 wake_up(&bc_serv
->sv_cb_waitq
);
360 r_xprt
->rx_stats
.bcall_count
++;
364 pr_warn("RPC/RDMA backchannel overflow\n");
365 xprt_disconnect_done(xprt
);
366 /* This receive buffer gets reposted automatically
367 * when the connection is re-established.
372 pr_warn("RPC/RDMA short backward direction call\n");
374 if (rpcrdma_ep_post_recv(&r_xprt
->rx_ia
, rep
))
375 xprt_disconnect_done(xprt
);
377 pr_warn("RPC: %s: reposting rep %p\n",