Merge tag 'nfsd-4.9' of git://linux-nfs.org/~bfields/linux
[cascardo/linux.git] / include / linux / sunrpc / xprt.h
1 /*
2  *  linux/include/linux/sunrpc/xprt.h
3  *
4  *  Declarations for the RPC transport interface.
5  *
6  *  Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
7  */
8
9 #ifndef _LINUX_SUNRPC_XPRT_H
10 #define _LINUX_SUNRPC_XPRT_H
11
12 #include <linux/uio.h>
13 #include <linux/socket.h>
14 #include <linux/in.h>
15 #include <linux/ktime.h>
16 #include <linux/kref.h>
17 #include <linux/sunrpc/sched.h>
18 #include <linux/sunrpc/xdr.h>
19 #include <linux/sunrpc/msg_prot.h>
20
21 #ifdef __KERNEL__
22
23 #define RPC_MIN_SLOT_TABLE      (2U)
24 #define RPC_DEF_SLOT_TABLE      (16U)
25 #define RPC_MAX_SLOT_TABLE_LIMIT        (65536U)
26 #define RPC_MAX_SLOT_TABLE      RPC_MAX_SLOT_TABLE_LIMIT
27
28 #define RPC_CWNDSHIFT           (8U)
29 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
30 #define RPC_INITCWND            RPC_CWNDSCALE
31 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
32 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
33
34 /*
35  * This describes a timeout strategy
36  */
37 struct rpc_timeout {
38         unsigned long           to_initval,             /* initial timeout */
39                                 to_maxval,              /* max timeout */
40                                 to_increment;           /* if !exponential */
41         unsigned int            to_retries;             /* max # of retries */
42         unsigned char           to_exponential;
43 };
44
45 enum rpc_display_format_t {
46         RPC_DISPLAY_ADDR = 0,
47         RPC_DISPLAY_PORT,
48         RPC_DISPLAY_PROTO,
49         RPC_DISPLAY_HEX_ADDR,
50         RPC_DISPLAY_HEX_PORT,
51         RPC_DISPLAY_NETID,
52         RPC_DISPLAY_MAX,
53 };
54
55 struct rpc_task;
56 struct rpc_xprt;
57 struct seq_file;
58 struct svc_serv;
59 struct net;
60
61 /*
62  * This describes a complete RPC request
63  */
64 struct rpc_rqst {
65         /*
66          * This is the user-visible part
67          */
68         struct rpc_xprt *       rq_xprt;                /* RPC client */
69         struct xdr_buf          rq_snd_buf;             /* send buffer */
70         struct xdr_buf          rq_rcv_buf;             /* recv buffer */
71
72         /*
73          * This is the private part
74          */
75         struct rpc_task *       rq_task;        /* RPC task data */
76         struct rpc_cred *       rq_cred;        /* Bound cred */
77         __be32                  rq_xid;         /* request XID */
78         int                     rq_cong;        /* has incremented xprt->cong */
79         u32                     rq_seqno;       /* gss seq no. used on req. */
80         int                     rq_enc_pages_num;
81         struct page             **rq_enc_pages; /* scratch pages for use by
82                                                    gss privacy code */
83         void (*rq_release_snd_buf)(struct rpc_rqst *); /* release rq_enc_pages */
84         struct list_head        rq_list;
85
86         __u32 *                 rq_buffer;      /* XDR encode buffer */
87         size_t                  rq_callsize,
88                                 rq_rcvsize;
89         size_t                  rq_xmit_bytes_sent;     /* total bytes sent */
90         size_t                  rq_reply_bytes_recvd;   /* total reply bytes */
91                                                         /* received */
92
93         struct xdr_buf          rq_private_buf;         /* The receive buffer
94                                                          * used in the softirq.
95                                                          */
96         unsigned long           rq_majortimeo;  /* major timeout alarm */
97         unsigned long           rq_timeout;     /* Current timeout value */
98         ktime_t                 rq_rtt;         /* round-trip time */
99         unsigned int            rq_retries;     /* # of retries */
100         unsigned int            rq_connect_cookie;
101                                                 /* A cookie used to track the
102                                                    state of the transport
103                                                    connection */
104         
105         /*
106          * Partial send handling
107          */
108         u32                     rq_bytes_sent;  /* Bytes we have sent */
109
110         ktime_t                 rq_xtime;       /* transmit time stamp */
111         int                     rq_ntrans;
112
113 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
114         struct list_head        rq_bc_list;     /* Callback service list */
115         unsigned long           rq_bc_pa_state; /* Backchannel prealloc state */
116         struct list_head        rq_bc_pa_list;  /* Backchannel prealloc list */
117 #endif /* CONFIG_SUNRPC_BACKCHANEL */
118 };
119 #define rq_svec                 rq_snd_buf.head
120 #define rq_slen                 rq_snd_buf.len
121
122 struct rpc_xprt_ops {
123         void            (*set_buffer_size)(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize);
124         int             (*reserve_xprt)(struct rpc_xprt *xprt, struct rpc_task *task);
125         void            (*release_xprt)(struct rpc_xprt *xprt, struct rpc_task *task);
126         void            (*alloc_slot)(struct rpc_xprt *xprt, struct rpc_task *task);
127         void            (*rpcbind)(struct rpc_task *task);
128         void            (*set_port)(struct rpc_xprt *xprt, unsigned short port);
129         void            (*connect)(struct rpc_xprt *xprt, struct rpc_task *task);
130         void *          (*buf_alloc)(struct rpc_task *task, size_t size);
131         void            (*buf_free)(void *buffer);
132         int             (*send_request)(struct rpc_task *task);
133         void            (*set_retrans_timeout)(struct rpc_task *task);
134         void            (*timer)(struct rpc_xprt *xprt, struct rpc_task *task);
135         void            (*release_request)(struct rpc_task *task);
136         void            (*close)(struct rpc_xprt *xprt);
137         void            (*destroy)(struct rpc_xprt *xprt);
138         void            (*print_stats)(struct rpc_xprt *xprt, struct seq_file *seq);
139         int             (*enable_swap)(struct rpc_xprt *xprt);
140         void            (*disable_swap)(struct rpc_xprt *xprt);
141         void            (*inject_disconnect)(struct rpc_xprt *xprt);
142         int             (*bc_setup)(struct rpc_xprt *xprt,
143                                     unsigned int min_reqs);
144         int             (*bc_up)(struct svc_serv *serv, struct net *net);
145         size_t          (*bc_maxpayload)(struct rpc_xprt *xprt);
146         void            (*bc_free_rqst)(struct rpc_rqst *rqst);
147         void            (*bc_destroy)(struct rpc_xprt *xprt,
148                                       unsigned int max_reqs);
149 };
150
151 /*
152  * RPC transport identifiers
153  *
154  * To preserve compatibility with the historical use of raw IP protocol
155  * id's for transport selection, UDP and TCP identifiers are specified
156  * with the previous values. No such restriction exists for new transports,
157  * except that they may not collide with these values (17 and 6,
158  * respectively).
159  */
160 #define XPRT_TRANSPORT_BC       (1 << 31)
161 enum xprt_transports {
162         XPRT_TRANSPORT_UDP      = IPPROTO_UDP,
163         XPRT_TRANSPORT_TCP      = IPPROTO_TCP,
164         XPRT_TRANSPORT_BC_TCP   = IPPROTO_TCP | XPRT_TRANSPORT_BC,
165         XPRT_TRANSPORT_RDMA     = 256,
166         XPRT_TRANSPORT_BC_RDMA  = XPRT_TRANSPORT_RDMA | XPRT_TRANSPORT_BC,
167         XPRT_TRANSPORT_LOCAL    = 257,
168 };
169
170 struct rpc_xprt {
171         struct kref             kref;           /* Reference count */
172         struct rpc_xprt_ops *   ops;            /* transport methods */
173
174         const struct rpc_timeout *timeout;      /* timeout parms */
175         struct sockaddr_storage addr;           /* server address */
176         size_t                  addrlen;        /* size of server address */
177         int                     prot;           /* IP protocol */
178
179         unsigned long           cong;           /* current congestion */
180         unsigned long           cwnd;           /* congestion window */
181
182         size_t                  max_payload;    /* largest RPC payload size,
183                                                    in bytes */
184         unsigned int            tsh_size;       /* size of transport specific
185                                                    header */
186
187         struct rpc_wait_queue   binding;        /* requests waiting on rpcbind */
188         struct rpc_wait_queue   sending;        /* requests waiting to send */
189         struct rpc_wait_queue   pending;        /* requests in flight */
190         struct rpc_wait_queue   backlog;        /* waiting for slot */
191         struct list_head        free;           /* free slots */
192         unsigned int            max_reqs;       /* max number of slots */
193         unsigned int            min_reqs;       /* min number of slots */
194         atomic_t                num_reqs;       /* total slots */
195         unsigned long           state;          /* transport state */
196         unsigned char           resvport   : 1; /* use a reserved port */
197         atomic_t                swapper;        /* we're swapping over this
198                                                    transport */
199         unsigned int            bind_index;     /* bind function index */
200
201         /*
202          * Multipath
203          */
204         struct list_head        xprt_switch;
205
206         /*
207          * Connection of transports
208          */
209         unsigned long           bind_timeout,
210                                 reestablish_timeout;
211         unsigned int            connect_cookie; /* A cookie that gets bumped
212                                                    every time the transport
213                                                    is reconnected */
214
215         /*
216          * Disconnection of idle transports
217          */
218         struct work_struct      task_cleanup;
219         struct timer_list       timer;
220         unsigned long           last_used,
221                                 idle_timeout,
222                                 max_reconnect_timeout;
223
224         /*
225          * Send stuff
226          */
227         spinlock_t              transport_lock; /* lock transport info */
228         spinlock_t              reserve_lock;   /* lock slot table */
229         u32                     xid;            /* Next XID value to use */
230         struct rpc_task *       snd_task;       /* Task blocked in send */
231         struct svc_xprt         *bc_xprt;       /* NFSv4.1 backchannel */
232 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
233         struct svc_serv         *bc_serv;       /* The RPC service which will */
234                                                 /* process the callback */
235         int                     bc_alloc_count; /* Total number of preallocs */
236         atomic_t                bc_free_slots;
237         spinlock_t              bc_pa_lock;     /* Protects the preallocated
238                                                  * items */
239         struct list_head        bc_pa_list;     /* List of preallocated
240                                                  * backchannel rpc_rqst's */
241 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
242         struct list_head        recv;
243
244         struct {
245                 unsigned long           bind_count,     /* total number of binds */
246                                         connect_count,  /* total number of connects */
247                                         connect_start,  /* connect start timestamp */
248                                         connect_time,   /* jiffies waiting for connect */
249                                         sends,          /* how many complete requests */
250                                         recvs,          /* how many complete requests */
251                                         bad_xids,       /* lookup_rqst didn't find XID */
252                                         max_slots;      /* max rpc_slots used */
253
254                 unsigned long long      req_u,          /* average requests on the wire */
255                                         bklog_u,        /* backlog queue utilization */
256                                         sending_u,      /* send q utilization */
257                                         pending_u;      /* pend q utilization */
258         } stat;
259
260         struct net              *xprt_net;
261         const char              *servername;
262         const char              *address_strings[RPC_DISPLAY_MAX];
263 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
264         struct dentry           *debugfs;               /* debugfs directory */
265         atomic_t                inject_disconnect;
266 #endif
267         struct rcu_head         rcu;
268 };
269
270 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
271 /*
272  * Backchannel flags
273  */
274 #define RPC_BC_PA_IN_USE        0x0001          /* Preallocated backchannel */
275                                                 /* buffer in use */
276 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
277
278 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
279 static inline int bc_prealloc(struct rpc_rqst *req)
280 {
281         return test_bit(RPC_BC_PA_IN_USE, &req->rq_bc_pa_state);
282 }
283 #else
284 static inline int bc_prealloc(struct rpc_rqst *req)
285 {
286         return 0;
287 }
288 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
289
290 #define XPRT_CREATE_INFINITE_SLOTS      (1U)
291 #define XPRT_CREATE_NO_IDLE_TIMEOUT     (1U << 1)
292
293 struct xprt_create {
294         int                     ident;          /* XPRT_TRANSPORT identifier */
295         struct net *            net;
296         struct sockaddr *       srcaddr;        /* optional local address */
297         struct sockaddr *       dstaddr;        /* remote peer address */
298         size_t                  addrlen;
299         const char              *servername;
300         struct svc_xprt         *bc_xprt;       /* NFSv4.1 backchannel */
301         struct rpc_xprt_switch  *bc_xps;
302         unsigned int            flags;
303 };
304
305 struct xprt_class {
306         struct list_head        list;
307         int                     ident;          /* XPRT_TRANSPORT identifier */
308         struct rpc_xprt *       (*setup)(struct xprt_create *);
309         struct module           *owner;
310         char                    name[32];
311 };
312
313 /*
314  * Generic internal transport functions
315  */
316 struct rpc_xprt         *xprt_create_transport(struct xprt_create *args);
317 void                    xprt_connect(struct rpc_task *task);
318 void                    xprt_reserve(struct rpc_task *task);
319 void                    xprt_retry_reserve(struct rpc_task *task);
320 int                     xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task);
321 int                     xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task);
322 void                    xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task);
323 void                    xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task);
324 bool                    xprt_prepare_transmit(struct rpc_task *task);
325 void                    xprt_transmit(struct rpc_task *task);
326 void                    xprt_end_transmit(struct rpc_task *task);
327 int                     xprt_adjust_timeout(struct rpc_rqst *req);
328 void                    xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task);
329 void                    xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task);
330 void                    xprt_release(struct rpc_task *task);
331 struct rpc_xprt *       xprt_get(struct rpc_xprt *xprt);
332 void                    xprt_put(struct rpc_xprt *xprt);
333 struct rpc_xprt *       xprt_alloc(struct net *net, size_t size,
334                                 unsigned int num_prealloc,
335                                 unsigned int max_req);
336 void                    xprt_free(struct rpc_xprt *);
337
338 static inline __be32 *xprt_skip_transport_header(struct rpc_xprt *xprt, __be32 *p)
339 {
340         return p + xprt->tsh_size;
341 }
342
343 static inline int
344 xprt_enable_swap(struct rpc_xprt *xprt)
345 {
346         return xprt->ops->enable_swap(xprt);
347 }
348
349 static inline void
350 xprt_disable_swap(struct rpc_xprt *xprt)
351 {
352         xprt->ops->disable_swap(xprt);
353 }
354
355 /*
356  * Transport switch helper functions
357  */
358 int                     xprt_register_transport(struct xprt_class *type);
359 int                     xprt_unregister_transport(struct xprt_class *type);
360 int                     xprt_load_transport(const char *);
361 void                    xprt_set_retrans_timeout_def(struct rpc_task *task);
362 void                    xprt_set_retrans_timeout_rtt(struct rpc_task *task);
363 void                    xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status);
364 void                    xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action);
365 void                    xprt_write_space(struct rpc_xprt *xprt);
366 void                    xprt_adjust_cwnd(struct rpc_xprt *xprt, struct rpc_task *task, int result);
367 struct rpc_rqst *       xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid);
368 void                    xprt_complete_rqst(struct rpc_task *task, int copied);
369 void                    xprt_release_rqst_cong(struct rpc_task *task);
370 void                    xprt_disconnect_done(struct rpc_xprt *xprt);
371 void                    xprt_force_disconnect(struct rpc_xprt *xprt);
372 void                    xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie);
373
374 bool                    xprt_lock_connect(struct rpc_xprt *, struct rpc_task *, void *);
375 void                    xprt_unlock_connect(struct rpc_xprt *, void *);
376
377 /*
378  * Reserved bit positions in xprt->state
379  */
380 #define XPRT_LOCKED             (0)
381 #define XPRT_CONNECTED          (1)
382 #define XPRT_CONNECTING         (2)
383 #define XPRT_CLOSE_WAIT         (3)
384 #define XPRT_BOUND              (4)
385 #define XPRT_BINDING            (5)
386 #define XPRT_CLOSING            (6)
387 #define XPRT_CONGESTED          (9)
388
389 static inline void xprt_set_connected(struct rpc_xprt *xprt)
390 {
391         set_bit(XPRT_CONNECTED, &xprt->state);
392 }
393
394 static inline void xprt_clear_connected(struct rpc_xprt *xprt)
395 {
396         clear_bit(XPRT_CONNECTED, &xprt->state);
397 }
398
399 static inline int xprt_connected(struct rpc_xprt *xprt)
400 {
401         return test_bit(XPRT_CONNECTED, &xprt->state);
402 }
403
404 static inline int xprt_test_and_set_connected(struct rpc_xprt *xprt)
405 {
406         return test_and_set_bit(XPRT_CONNECTED, &xprt->state);
407 }
408
409 static inline int xprt_test_and_clear_connected(struct rpc_xprt *xprt)
410 {
411         return test_and_clear_bit(XPRT_CONNECTED, &xprt->state);
412 }
413
414 static inline void xprt_clear_connecting(struct rpc_xprt *xprt)
415 {
416         smp_mb__before_atomic();
417         clear_bit(XPRT_CONNECTING, &xprt->state);
418         smp_mb__after_atomic();
419 }
420
421 static inline int xprt_connecting(struct rpc_xprt *xprt)
422 {
423         return test_bit(XPRT_CONNECTING, &xprt->state);
424 }
425
426 static inline int xprt_test_and_set_connecting(struct rpc_xprt *xprt)
427 {
428         return test_and_set_bit(XPRT_CONNECTING, &xprt->state);
429 }
430
431 static inline void xprt_set_bound(struct rpc_xprt *xprt)
432 {
433         test_and_set_bit(XPRT_BOUND, &xprt->state);
434 }
435
436 static inline int xprt_bound(struct rpc_xprt *xprt)
437 {
438         return test_bit(XPRT_BOUND, &xprt->state);
439 }
440
441 static inline void xprt_clear_bound(struct rpc_xprt *xprt)
442 {
443         clear_bit(XPRT_BOUND, &xprt->state);
444 }
445
446 static inline void xprt_clear_binding(struct rpc_xprt *xprt)
447 {
448         smp_mb__before_atomic();
449         clear_bit(XPRT_BINDING, &xprt->state);
450         smp_mb__after_atomic();
451 }
452
453 static inline int xprt_test_and_set_binding(struct rpc_xprt *xprt)
454 {
455         return test_and_set_bit(XPRT_BINDING, &xprt->state);
456 }
457
458 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
459 extern unsigned int rpc_inject_disconnect;
460 static inline void xprt_inject_disconnect(struct rpc_xprt *xprt)
461 {
462         if (!rpc_inject_disconnect)
463                 return;
464         if (atomic_dec_return(&xprt->inject_disconnect))
465                 return;
466         atomic_set(&xprt->inject_disconnect, rpc_inject_disconnect);
467         xprt->ops->inject_disconnect(xprt);
468 }
469 #else
470 static inline void xprt_inject_disconnect(struct rpc_xprt *xprt)
471 {
472 }
473 #endif
474
475 #endif /* __KERNEL__*/
476
477 #endif /* _LINUX_SUNRPC_XPRT_H */