network - Tokenize NFS, fix MP races
[dragonfly.git] / sys / vfs / nfs / nfs_iod.c
1 /*
2  * Copyright (c) 2009 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 /*
35  * NFSIOD operations - now built into the kernel.
36  */
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/proc.h>
40 #include <sys/malloc.h>
41 #include <sys/mount.h>
42 #include <sys/kernel.h>
43 #include <sys/mbuf.h>
44 #include <sys/vnode.h>
45 #include <sys/fcntl.h>
46 #include <sys/protosw.h>
47 #include <sys/resourcevar.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/socketops.h>
51 #include <sys/syslog.h>
52 #include <sys/thread.h>
53 #include <sys/tprintf.h>
54 #include <sys/sysctl.h>
55 #include <sys/signalvar.h>
56 #include <sys/mutex.h>
57
58 #include <sys/signal2.h>
59 #include <sys/thread2.h>
60 #include <sys/mutex2.h>
61 #include <sys/mplock2.h>
62
63 #include <netinet/in.h>
64 #include <netinet/tcp.h>
65
66 #include "rpcv2.h"
67 #include "nfsproto.h"
68 #include "nfs.h"
69 #include "xdr_subs.h"
70 #include "nfsm_subs.h"
71 #include "nfsmount.h"
72 #include "nfsnode.h"
73 #include "nfsrtt.h"
74
75 /*
76  * nfs service connection reader thread
77  */
78 void
79 nfssvc_iod_reader(void *arg)
80 {
81         struct nfsmount *nmp = arg;
82         struct nfsm_info *info;
83         struct nfsreq *req;
84         int error;
85
86         lwkt_gettoken(&nmp->nm_token);
87
88         if (nmp->nm_rxstate == NFSSVC_INIT)
89                 nmp->nm_rxstate = NFSSVC_PENDING;
90         for (;;) {
91                 if (nmp->nm_rxstate == NFSSVC_WAITING) {
92                         if (TAILQ_FIRST(&nmp->nm_reqq) == NULL &&
93                             TAILQ_FIRST(&nmp->nm_reqrxq) == NULL) {
94                                 tsleep(&nmp->nm_rxstate, 0, "nfsidl", 0);
95                         } else {
96                                 /*
97                                  * This can happen during shutdown, we don't
98                                  * want to hardloop.
99                                  */
100                                 error = nfs_reply(nmp, NULL);
101                                 if (error && error != EWOULDBLOCK) {
102                                         tsleep(&nmp->nm_rxstate, 0,
103                                                 "nfsxxx", hz / 10);
104                                 }
105                         }
106                         continue;
107                 }
108                 if (nmp->nm_rxstate != NFSSVC_PENDING)
109                         break;
110                 nmp->nm_rxstate = NFSSVC_WAITING;
111
112                 /*
113                  * Process requests which have received replies.  Only
114                  * process the post-reply states.  If we get EINPROGRESS
115                  * it means the request went back to an auth or retransmit
116                  * state and we let the iod_writer thread deal with it.
117                  *
118                  * Any lock on the request is strictly temporary due to
119                  * MP races (XXX).
120                  *
121                  * If the request completes we run the info->done call
122                  * to finish up the I/O.
123                  */
124                 while ((req = TAILQ_FIRST(&nmp->nm_reqrxq)) != NULL) {
125                         if (req->r_flags & R_LOCKED) {
126                                 while (req->r_flags & R_LOCKED) {
127                                         req->r_flags |= R_WANTED;
128                                         tsleep(req, 0, "nfstrac", 0);
129                                 }
130                                 continue;
131                         }
132                         TAILQ_REMOVE(&nmp->nm_reqrxq, req, r_chain);
133                         info = req->r_info;
134                         KKASSERT(info);
135                         info->error = nfs_request(info,
136                                                   NFSM_STATE_PROCESSREPLY,
137                                                   NFSM_STATE_DONE);
138                         if (info->error == EINPROGRESS) {
139                                 kprintf("rxq: move info %p back to txq\n", info);
140                                 TAILQ_INSERT_TAIL(&nmp->nm_reqtxq, req, r_chain);
141                                 nfssvc_iod_writer_wakeup(nmp);
142                         } else {
143                                 atomic_subtract_int(&nmp->nm_bioqlen, 1);
144                                 info->done(info);
145                         }
146                 }
147         }
148         nmp->nm_rxthread = NULL;
149         nmp->nm_rxstate = NFSSVC_DONE;
150
151         lwkt_reltoken(&nmp->nm_token);
152         wakeup(&nmp->nm_rxthread);
153 }
154
155 /*
156  * nfs service connection writer thread
157  *
158  * The writer sits on the send side of the client's socket and
159  * does both the initial processing of BIOs and also transmission
160  * and retransmission of nfsreq's.
161  *
162  * The writer processes both new BIOs from nm_bioq and retransmit
163  * or state machine jumpbacks from nm_reqtxq
164  */
165 void
166 nfssvc_iod_writer(void *arg)
167 {
168         struct nfsmount *nmp = arg;
169         struct bio *bio;
170         struct nfsreq *req;
171         struct vnode *vp;
172         nfsm_info_t info;
173
174         lwkt_gettoken(&nmp->nm_token);
175
176         if (nmp->nm_txstate == NFSSVC_INIT)
177                 nmp->nm_txstate = NFSSVC_PENDING;
178
179         for (;;) {
180                 if (nmp->nm_txstate == NFSSVC_WAITING) {
181                         tsleep(&nmp->nm_txstate, 0, "nfsidl", 0);
182                         continue;
183                 }
184                 if (nmp->nm_txstate != NFSSVC_PENDING)
185                         break;
186                 nmp->nm_txstate = NFSSVC_WAITING;
187
188                 /*
189                  * Eep, we could blow out the mbuf allocator if we just
190                  * did everything the kernel wanted us to do.
191                  */
192                 while ((bio = TAILQ_FIRST(&nmp->nm_bioq)) != NULL) {
193                         if (nmp->nm_reqqlen > nfs_maxasyncbio)
194                                 break;
195                         TAILQ_REMOVE(&nmp->nm_bioq, bio, bio_act);
196                         vp = bio->bio_driver_info;
197                         nfs_startio(vp, bio, NULL);
198                 }
199
200                 /*
201                  * Process reauths & retransmits.  If we get an EINPROGRESS
202                  * it means the state transitioned to WAITREPLY or later.
203                  * Otherwise the request completed (probably with an error
204                  * since we didn't get to a replied state).
205                  */
206                 while ((req = TAILQ_FIRST(&nmp->nm_reqtxq)) != NULL) {
207                         TAILQ_REMOVE(&nmp->nm_reqtxq, req, r_chain);
208                         info = req->r_info;
209                         KKASSERT(info);
210                         info->error = nfs_request(info,
211                                                   NFSM_STATE_AUTH,
212                                                   NFSM_STATE_WAITREPLY);
213                         if (info->error == EINPROGRESS) {
214                                 ;
215                         } else {
216                                 atomic_subtract_int(&nmp->nm_bioqlen, 1);
217                                 info->done(info);
218                         }
219                 }
220         }
221         nmp->nm_txthread = NULL;
222         nmp->nm_txstate = NFSSVC_DONE;
223         lwkt_reltoken(&nmp->nm_token);
224         wakeup(&nmp->nm_txthread);
225 }
226
227 void
228 nfssvc_iod_stop1(struct nfsmount *nmp)
229 {
230         nmp->nm_txstate = NFSSVC_STOPPING;
231         nmp->nm_rxstate = NFSSVC_STOPPING;
232 }
233
234 void
235 nfssvc_iod_stop2(struct nfsmount *nmp)
236 {
237         wakeup(&nmp->nm_txstate);
238         while (nmp->nm_txthread)
239                 tsleep(&nmp->nm_txthread, 0, "nfssttx", hz*2);
240         wakeup(&nmp->nm_rxstate);
241         while (nmp->nm_rxthread)
242                 tsleep(&nmp->nm_rxthread, 0, "nfsstrx", hz*2);
243 }
244
245 void
246 nfssvc_iod_writer_wakeup(struct nfsmount *nmp)
247 {
248         if (nmp->nm_txstate == NFSSVC_WAITING) {
249                 nmp->nm_txstate = NFSSVC_PENDING;
250                 wakeup(&nmp->nm_txstate);
251         }
252 }
253
254 void
255 nfssvc_iod_reader_wakeup(struct nfsmount *nmp)
256 {
257         if (nmp->nm_rxstate == NFSSVC_WAITING) {
258                 nmp->nm_rxstate = NFSSVC_PENDING;
259                 wakeup(&nmp->nm_rxstate);
260         }
261 }