| 1 | /* |
| 2 | * linux/net/sunrpc/xprt.c |
| 3 | * |
| 4 | * This is a generic RPC call interface supporting congestion avoidance, |
| 5 | * and asynchronous calls. |
| 6 | * |
| 7 | * The interface works like this: |
| 8 | * |
| 9 | * - When a process places a call, it allocates a request slot if |
| 10 | * one is available. Otherwise, it sleeps on the backlog queue |
| 11 | * (xprt_reserve). |
| 12 | * - Next, the caller puts together the RPC message, stuffs it into |
| 13 | * the request struct, and calls xprt_transmit(). |
| 14 | * - xprt_transmit sends the message and installs the caller on the |
| 15 | * transport's wait list. At the same time, if a reply is expected, |
| 16 | * it installs a timer that is run after the packet's timeout has |
| 17 | * expired. |
| 18 | * - When a packet arrives, the data_ready handler walks the list of |
| 19 | * pending requests for that transport. If a matching XID is found, the |
| 20 | * caller is woken up, and the timer removed. |
| 21 | * - When no reply arrives within the timeout interval, the timer is |
| 22 | * fired by the kernel and runs xprt_timer(). It either adjusts the |
| 23 | * timeout values (minor timeout) or wakes up the caller with a status |
| 24 | * of -ETIMEDOUT. |
| 25 | * - When the caller receives a notification from RPC that a reply arrived, |
| 26 | * it should release the RPC slot, and process the reply. |
| 27 | * If the call timed out, it may choose to retry the operation by |
| 28 | * adjusting the initial timeout value, and simply calling rpc_call |
| 29 | * again. |
| 30 | * |
| 31 | * Support for async RPC is done through a set of RPC-specific scheduling |
| 32 | * primitives that `transparently' work for processes as well as async |
| 33 | * tasks that rely on callbacks. |
| 34 | * |
| 35 | * Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de> |
| 36 | * |
| 37 | * Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com> |
| 38 | */ |
| 39 | |
| 40 | #include <linux/module.h> |
| 41 | |
| 42 | #include <linux/types.h> |
| 43 | #include <linux/interrupt.h> |
| 44 | #include <linux/workqueue.h> |
| 45 | #include <linux/net.h> |
| 46 | #include <linux/ktime.h> |
| 47 | |
| 48 | #include <linux/sunrpc/clnt.h> |
| 49 | #include <linux/sunrpc/metrics.h> |
| 50 | #include <linux/sunrpc/bc_xprt.h> |
| 51 | #include <linux/rcupdate.h> |
| 52 | |
| 53 | #include <trace/events/sunrpc.h> |
| 54 | |
| 55 | #include "sunrpc.h" |
| 56 | |
| 57 | /* |
| 58 | * Local variables |
| 59 | */ |
| 60 | |
| 61 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
| 62 | # define RPCDBG_FACILITY RPCDBG_XPRT |
| 63 | #endif |
| 64 | |
| 65 | /* |
| 66 | * Local functions |
| 67 | */ |
| 68 | static void xprt_init(struct rpc_xprt *xprt, struct net *net); |
| 69 | static void xprt_request_init(struct rpc_task *, struct rpc_xprt *); |
| 70 | static void xprt_connect_status(struct rpc_task *task); |
| 71 | static int __xprt_get_cong(struct rpc_xprt *, struct rpc_task *); |
| 72 | static void __xprt_put_cong(struct rpc_xprt *, struct rpc_rqst *); |
| 73 | static void xprt_destroy(struct rpc_xprt *xprt); |
| 74 | |
| 75 | static DEFINE_SPINLOCK(xprt_list_lock); |
| 76 | static LIST_HEAD(xprt_list); |
| 77 | |
| 78 | /** |
| 79 | * xprt_register_transport - register a transport implementation |
| 80 | * @transport: transport to register |
| 81 | * |
| 82 | * If a transport implementation is loaded as a kernel module, it can |
| 83 | * call this interface to make itself known to the RPC client. |
| 84 | * |
| 85 | * Returns: |
| 86 | * 0: transport successfully registered |
| 87 | * -EEXIST: transport already registered |
| 88 | * -EINVAL: transport module being unloaded |
| 89 | */ |
| 90 | int xprt_register_transport(struct xprt_class *transport) |
| 91 | { |
| 92 | struct xprt_class *t; |
| 93 | int result; |
| 94 | |
| 95 | result = -EEXIST; |
| 96 | spin_lock(&xprt_list_lock); |
| 97 | list_for_each_entry(t, &xprt_list, list) { |
| 98 | /* don't register the same transport class twice */ |
| 99 | if (t->ident == transport->ident) |
| 100 | goto out; |
| 101 | } |
| 102 | |
| 103 | list_add_tail(&transport->list, &xprt_list); |
| 104 | printk(KERN_INFO "RPC: Registered %s transport module.\n", |
| 105 | transport->name); |
| 106 | result = 0; |
| 107 | |
| 108 | out: |
| 109 | spin_unlock(&xprt_list_lock); |
| 110 | return result; |
| 111 | } |
| 112 | EXPORT_SYMBOL_GPL(xprt_register_transport); |
| 113 | |
| 114 | /** |
| 115 | * xprt_unregister_transport - unregister a transport implementation |
| 116 | * @transport: transport to unregister |
| 117 | * |
| 118 | * Returns: |
| 119 | * 0: transport successfully unregistered |
| 120 | * -ENOENT: transport never registered |
| 121 | */ |
| 122 | int xprt_unregister_transport(struct xprt_class *transport) |
| 123 | { |
| 124 | struct xprt_class *t; |
| 125 | int result; |
| 126 | |
| 127 | result = 0; |
| 128 | spin_lock(&xprt_list_lock); |
| 129 | list_for_each_entry(t, &xprt_list, list) { |
| 130 | if (t == transport) { |
| 131 | printk(KERN_INFO |
| 132 | "RPC: Unregistered %s transport module.\n", |
| 133 | transport->name); |
| 134 | list_del_init(&transport->list); |
| 135 | goto out; |
| 136 | } |
| 137 | } |
| 138 | result = -ENOENT; |
| 139 | |
| 140 | out: |
| 141 | spin_unlock(&xprt_list_lock); |
| 142 | return result; |
| 143 | } |
| 144 | EXPORT_SYMBOL_GPL(xprt_unregister_transport); |
| 145 | |
| 146 | /** |
| 147 | * xprt_load_transport - load a transport implementation |
| 148 | * @transport_name: transport to load |
| 149 | * |
| 150 | * Returns: |
| 151 | * 0: transport successfully loaded |
| 152 | * -ENOENT: transport module not available |
| 153 | */ |
| 154 | int xprt_load_transport(const char *transport_name) |
| 155 | { |
| 156 | struct xprt_class *t; |
| 157 | int result; |
| 158 | |
| 159 | result = 0; |
| 160 | spin_lock(&xprt_list_lock); |
| 161 | list_for_each_entry(t, &xprt_list, list) { |
| 162 | if (strcmp(t->name, transport_name) == 0) { |
| 163 | spin_unlock(&xprt_list_lock); |
| 164 | goto out; |
| 165 | } |
| 166 | } |
| 167 | spin_unlock(&xprt_list_lock); |
| 168 | result = request_module("xprt%s", transport_name); |
| 169 | out: |
| 170 | return result; |
| 171 | } |
| 172 | EXPORT_SYMBOL_GPL(xprt_load_transport); |
| 173 | |
| 174 | /** |
| 175 | * xprt_reserve_xprt - serialize write access to transports |
| 176 | * @task: task that is requesting access to the transport |
| 177 | * @xprt: pointer to the target transport |
| 178 | * |
| 179 | * This prevents mixing the payload of separate requests, and prevents |
| 180 | * transport connects from colliding with writes. No congestion control |
| 181 | * is provided. |
| 182 | */ |
| 183 | int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task) |
| 184 | { |
| 185 | struct rpc_rqst *req = task->tk_rqstp; |
| 186 | int priority; |
| 187 | |
| 188 | if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) { |
| 189 | if (task == xprt->snd_task) |
| 190 | return 1; |
| 191 | goto out_sleep; |
| 192 | } |
| 193 | xprt->snd_task = task; |
| 194 | if (req != NULL) |
| 195 | req->rq_ntrans++; |
| 196 | |
| 197 | return 1; |
| 198 | |
| 199 | out_sleep: |
| 200 | dprintk("RPC: %5u failed to lock transport %p\n", |
| 201 | task->tk_pid, xprt); |
| 202 | task->tk_timeout = 0; |
| 203 | task->tk_status = -EAGAIN; |
| 204 | if (req == NULL) |
| 205 | priority = RPC_PRIORITY_LOW; |
| 206 | else if (!req->rq_ntrans) |
| 207 | priority = RPC_PRIORITY_NORMAL; |
| 208 | else |
| 209 | priority = RPC_PRIORITY_HIGH; |
| 210 | rpc_sleep_on_priority(&xprt->sending, task, NULL, priority); |
| 211 | return 0; |
| 212 | } |
| 213 | EXPORT_SYMBOL_GPL(xprt_reserve_xprt); |
| 214 | |
| 215 | static void xprt_clear_locked(struct rpc_xprt *xprt) |
| 216 | { |
| 217 | xprt->snd_task = NULL; |
| 218 | if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state)) { |
| 219 | smp_mb__before_atomic(); |
| 220 | clear_bit(XPRT_LOCKED, &xprt->state); |
| 221 | smp_mb__after_atomic(); |
| 222 | } else |
| 223 | queue_work(xprtiod_workqueue, &xprt->task_cleanup); |
| 224 | } |
| 225 | |
| 226 | /* |
| 227 | * xprt_reserve_xprt_cong - serialize write access to transports |
| 228 | * @task: task that is requesting access to the transport |
| 229 | * |
| 230 | * Same as xprt_reserve_xprt, but Van Jacobson congestion control is |
| 231 | * integrated into the decision of whether a request is allowed to be |
| 232 | * woken up and given access to the transport. |
| 233 | */ |
| 234 | int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task) |
| 235 | { |
| 236 | struct rpc_rqst *req = task->tk_rqstp; |
| 237 | int priority; |
| 238 | |
| 239 | if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) { |
| 240 | if (task == xprt->snd_task) |
| 241 | return 1; |
| 242 | goto out_sleep; |
| 243 | } |
| 244 | if (req == NULL) { |
| 245 | xprt->snd_task = task; |
| 246 | return 1; |
| 247 | } |
| 248 | if (__xprt_get_cong(xprt, task)) { |
| 249 | xprt->snd_task = task; |
| 250 | req->rq_ntrans++; |
| 251 | return 1; |
| 252 | } |
| 253 | xprt_clear_locked(xprt); |
| 254 | out_sleep: |
| 255 | if (req) |
| 256 | __xprt_put_cong(xprt, req); |
| 257 | dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt); |
| 258 | task->tk_timeout = 0; |
| 259 | task->tk_status = -EAGAIN; |
| 260 | if (req == NULL) |
| 261 | priority = RPC_PRIORITY_LOW; |
| 262 | else if (!req->rq_ntrans) |
| 263 | priority = RPC_PRIORITY_NORMAL; |
| 264 | else |
| 265 | priority = RPC_PRIORITY_HIGH; |
| 266 | rpc_sleep_on_priority(&xprt->sending, task, NULL, priority); |
| 267 | return 0; |
| 268 | } |
| 269 | EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong); |
| 270 | |
| 271 | static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task) |
| 272 | { |
| 273 | int retval; |
| 274 | |
| 275 | spin_lock_bh(&xprt->transport_lock); |
| 276 | retval = xprt->ops->reserve_xprt(xprt, task); |
| 277 | spin_unlock_bh(&xprt->transport_lock); |
| 278 | return retval; |
| 279 | } |
| 280 | |
| 281 | static bool __xprt_lock_write_func(struct rpc_task *task, void *data) |
| 282 | { |
| 283 | struct rpc_xprt *xprt = data; |
| 284 | struct rpc_rqst *req; |
| 285 | |
| 286 | req = task->tk_rqstp; |
| 287 | xprt->snd_task = task; |
| 288 | if (req) |
| 289 | req->rq_ntrans++; |
| 290 | return true; |
| 291 | } |
| 292 | |
| 293 | static void __xprt_lock_write_next(struct rpc_xprt *xprt) |
| 294 | { |
| 295 | if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) |
| 296 | return; |
| 297 | |
| 298 | if (rpc_wake_up_first_on_wq(xprtiod_workqueue, &xprt->sending, |
| 299 | __xprt_lock_write_func, xprt)) |
| 300 | return; |
| 301 | xprt_clear_locked(xprt); |
| 302 | } |
| 303 | |
| 304 | static bool __xprt_lock_write_cong_func(struct rpc_task *task, void *data) |
| 305 | { |
| 306 | struct rpc_xprt *xprt = data; |
| 307 | struct rpc_rqst *req; |
| 308 | |
| 309 | req = task->tk_rqstp; |
| 310 | if (req == NULL) { |
| 311 | xprt->snd_task = task; |
| 312 | return true; |
| 313 | } |
| 314 | if (__xprt_get_cong(xprt, task)) { |
| 315 | xprt->snd_task = task; |
| 316 | req->rq_ntrans++; |
| 317 | return true; |
| 318 | } |
| 319 | return false; |
| 320 | } |
| 321 | |
| 322 | static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt) |
| 323 | { |
| 324 | if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) |
| 325 | return; |
| 326 | if (RPCXPRT_CONGESTED(xprt)) |
| 327 | goto out_unlock; |
| 328 | if (rpc_wake_up_first_on_wq(xprtiod_workqueue, &xprt->sending, |
| 329 | __xprt_lock_write_cong_func, xprt)) |
| 330 | return; |
| 331 | out_unlock: |
| 332 | xprt_clear_locked(xprt); |
| 333 | } |
| 334 | |
| 335 | static void xprt_task_clear_bytes_sent(struct rpc_task *task) |
| 336 | { |
| 337 | if (task != NULL) { |
| 338 | struct rpc_rqst *req = task->tk_rqstp; |
| 339 | if (req != NULL) |
| 340 | req->rq_bytes_sent = 0; |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | /** |
| 345 | * xprt_release_xprt - allow other requests to use a transport |
| 346 | * @xprt: transport with other tasks potentially waiting |
| 347 | * @task: task that is releasing access to the transport |
| 348 | * |
| 349 | * Note that "task" can be NULL. No congestion control is provided. |
| 350 | */ |
| 351 | void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task) |
| 352 | { |
| 353 | if (xprt->snd_task == task) { |
| 354 | xprt_task_clear_bytes_sent(task); |
| 355 | xprt_clear_locked(xprt); |
| 356 | __xprt_lock_write_next(xprt); |
| 357 | } |
| 358 | } |
| 359 | EXPORT_SYMBOL_GPL(xprt_release_xprt); |
| 360 | |
| 361 | /** |
| 362 | * xprt_release_xprt_cong - allow other requests to use a transport |
| 363 | * @xprt: transport with other tasks potentially waiting |
| 364 | * @task: task that is releasing access to the transport |
| 365 | * |
| 366 | * Note that "task" can be NULL. Another task is awoken to use the |
| 367 | * transport if the transport's congestion window allows it. |
| 368 | */ |
| 369 | void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task) |
| 370 | { |
| 371 | if (xprt->snd_task == task) { |
| 372 | xprt_task_clear_bytes_sent(task); |
| 373 | xprt_clear_locked(xprt); |
| 374 | __xprt_lock_write_next_cong(xprt); |
| 375 | } |
| 376 | } |
| 377 | EXPORT_SYMBOL_GPL(xprt_release_xprt_cong); |
| 378 | |
| 379 | static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task) |
| 380 | { |
| 381 | spin_lock_bh(&xprt->transport_lock); |
| 382 | xprt->ops->release_xprt(xprt, task); |
| 383 | spin_unlock_bh(&xprt->transport_lock); |
| 384 | } |
| 385 | |
| 386 | /* |
| 387 | * Van Jacobson congestion avoidance. Check if the congestion window |
| 388 | * overflowed. Put the task to sleep if this is the case. |
| 389 | */ |
| 390 | static int |
| 391 | __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task) |
| 392 | { |
| 393 | struct rpc_rqst *req = task->tk_rqstp; |
| 394 | |
| 395 | if (req->rq_cong) |
| 396 | return 1; |
| 397 | dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n", |
| 398 | task->tk_pid, xprt->cong, xprt->cwnd); |
| 399 | if (RPCXPRT_CONGESTED(xprt)) |
| 400 | return 0; |
| 401 | req->rq_cong = 1; |
| 402 | xprt->cong += RPC_CWNDSCALE; |
| 403 | return 1; |
| 404 | } |
| 405 | |
| 406 | /* |
| 407 | * Adjust the congestion window, and wake up the next task |
| 408 | * that has been sleeping due to congestion |
| 409 | */ |
| 410 | static void |
| 411 | __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req) |
| 412 | { |
| 413 | if (!req->rq_cong) |
| 414 | return; |
| 415 | req->rq_cong = 0; |
| 416 | xprt->cong -= RPC_CWNDSCALE; |
| 417 | __xprt_lock_write_next_cong(xprt); |
| 418 | } |
| 419 | |
| 420 | /** |
| 421 | * xprt_release_rqst_cong - housekeeping when request is complete |
| 422 | * @task: RPC request that recently completed |
| 423 | * |
| 424 | * Useful for transports that require congestion control. |
| 425 | */ |
| 426 | void xprt_release_rqst_cong(struct rpc_task *task) |
| 427 | { |
| 428 | struct rpc_rqst *req = task->tk_rqstp; |
| 429 | |
| 430 | __xprt_put_cong(req->rq_xprt, req); |
| 431 | } |
| 432 | EXPORT_SYMBOL_GPL(xprt_release_rqst_cong); |
| 433 | |
| 434 | /** |
| 435 | * xprt_adjust_cwnd - adjust transport congestion window |
| 436 | * @xprt: pointer to xprt |
| 437 | * @task: recently completed RPC request used to adjust window |
| 438 | * @result: result code of completed RPC request |
| 439 | * |
| 440 | * The transport code maintains an estimate on the maximum number of out- |
| 441 | * standing RPC requests, using a smoothed version of the congestion |
| 442 | * avoidance implemented in 44BSD. This is basically the Van Jacobson |
| 443 | * congestion algorithm: If a retransmit occurs, the congestion window is |
| 444 | * halved; otherwise, it is incremented by 1/cwnd when |
| 445 | * |
| 446 | * - a reply is received and |
| 447 | * - a full number of requests are outstanding and |
| 448 | * - the congestion window hasn't been updated recently. |
| 449 | */ |
| 450 | void xprt_adjust_cwnd(struct rpc_xprt *xprt, struct rpc_task *task, int result) |
| 451 | { |
| 452 | struct rpc_rqst *req = task->tk_rqstp; |
| 453 | unsigned long cwnd = xprt->cwnd; |
| 454 | |
| 455 | if (result >= 0 && cwnd <= xprt->cong) { |
| 456 | /* The (cwnd >> 1) term makes sure |
| 457 | * the result gets rounded properly. */ |
| 458 | cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd; |
| 459 | if (cwnd > RPC_MAXCWND(xprt)) |
| 460 | cwnd = RPC_MAXCWND(xprt); |
| 461 | __xprt_lock_write_next_cong(xprt); |
| 462 | } else if (result == -ETIMEDOUT) { |
| 463 | cwnd >>= 1; |
| 464 | if (cwnd < RPC_CWNDSCALE) |
| 465 | cwnd = RPC_CWNDSCALE; |
| 466 | } |
| 467 | dprintk("RPC: cong %ld, cwnd was %ld, now %ld\n", |
| 468 | xprt->cong, xprt->cwnd, cwnd); |
| 469 | xprt->cwnd = cwnd; |
| 470 | __xprt_put_cong(xprt, req); |
| 471 | } |
| 472 | EXPORT_SYMBOL_GPL(xprt_adjust_cwnd); |
| 473 | |
| 474 | /** |
| 475 | * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue |
| 476 | * @xprt: transport with waiting tasks |
| 477 | * @status: result code to plant in each task before waking it |
| 478 | * |
| 479 | */ |
| 480 | void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status) |
| 481 | { |
| 482 | if (status < 0) |
| 483 | rpc_wake_up_status(&xprt->pending, status); |
| 484 | else |
| 485 | rpc_wake_up(&xprt->pending); |
| 486 | } |
| 487 | EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks); |
| 488 | |
| 489 | /** |
| 490 | * xprt_wait_for_buffer_space - wait for transport output buffer to clear |
| 491 | * @task: task to be put to sleep |
| 492 | * @action: function pointer to be executed after wait |
| 493 | * |
| 494 | * Note that we only set the timer for the case of RPC_IS_SOFT(), since |
| 495 | * we don't in general want to force a socket disconnection due to |
| 496 | * an incomplete RPC call transmission. |
| 497 | */ |
| 498 | void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action) |
| 499 | { |
| 500 | struct rpc_rqst *req = task->tk_rqstp; |
| 501 | struct rpc_xprt *xprt = req->rq_xprt; |
| 502 | |
| 503 | task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0; |
| 504 | rpc_sleep_on(&xprt->pending, task, action); |
| 505 | } |
| 506 | EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space); |
| 507 | |
| 508 | /** |
| 509 | * xprt_write_space - wake the task waiting for transport output buffer space |
| 510 | * @xprt: transport with waiting tasks |
| 511 | * |
| 512 | * Can be called in a soft IRQ context, so xprt_write_space never sleeps. |
| 513 | */ |
| 514 | void xprt_write_space(struct rpc_xprt *xprt) |
| 515 | { |
| 516 | spin_lock_bh(&xprt->transport_lock); |
| 517 | if (xprt->snd_task) { |
| 518 | dprintk("RPC: write space: waking waiting task on " |
| 519 | "xprt %p\n", xprt); |
| 520 | rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task); |
| 521 | } |
| 522 | spin_unlock_bh(&xprt->transport_lock); |
| 523 | } |
| 524 | EXPORT_SYMBOL_GPL(xprt_write_space); |
| 525 | |
| 526 | /** |
| 527 | * xprt_set_retrans_timeout_def - set a request's retransmit timeout |
| 528 | * @task: task whose timeout is to be set |
| 529 | * |
| 530 | * Set a request's retransmit timeout based on the transport's |
| 531 | * default timeout parameters. Used by transports that don't adjust |
| 532 | * the retransmit timeout based on round-trip time estimation. |
| 533 | */ |
| 534 | void xprt_set_retrans_timeout_def(struct rpc_task *task) |
| 535 | { |
| 536 | task->tk_timeout = task->tk_rqstp->rq_timeout; |
| 537 | } |
| 538 | EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def); |
| 539 | |
| 540 | /** |
| 541 | * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout |
| 542 | * @task: task whose timeout is to be set |
| 543 | * |
| 544 | * Set a request's retransmit timeout using the RTT estimator. |
| 545 | */ |
| 546 | void xprt_set_retrans_timeout_rtt(struct rpc_task *task) |
| 547 | { |
| 548 | int timer = task->tk_msg.rpc_proc->p_timer; |
| 549 | struct rpc_clnt *clnt = task->tk_client; |
| 550 | struct rpc_rtt *rtt = clnt->cl_rtt; |
| 551 | struct rpc_rqst *req = task->tk_rqstp; |
| 552 | unsigned long max_timeout = clnt->cl_timeout->to_maxval; |
| 553 | |
| 554 | task->tk_timeout = rpc_calc_rto(rtt, timer); |
| 555 | task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries; |
| 556 | if (task->tk_timeout > max_timeout || task->tk_timeout == 0) |
| 557 | task->tk_timeout = max_timeout; |
| 558 | } |
| 559 | EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt); |
| 560 | |
| 561 | static void xprt_reset_majortimeo(struct rpc_rqst *req) |
| 562 | { |
| 563 | const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout; |
| 564 | |
| 565 | req->rq_majortimeo = req->rq_timeout; |
| 566 | if (to->to_exponential) |
| 567 | req->rq_majortimeo <<= to->to_retries; |
| 568 | else |
| 569 | req->rq_majortimeo += to->to_increment * to->to_retries; |
| 570 | if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0) |
| 571 | req->rq_majortimeo = to->to_maxval; |
| 572 | req->rq_majortimeo += jiffies; |
| 573 | } |
| 574 | |
| 575 | /** |
| 576 | * xprt_adjust_timeout - adjust timeout values for next retransmit |
| 577 | * @req: RPC request containing parameters to use for the adjustment |
| 578 | * |
| 579 | */ |
| 580 | int xprt_adjust_timeout(struct rpc_rqst *req) |
| 581 | { |
| 582 | struct rpc_xprt *xprt = req->rq_xprt; |
| 583 | const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout; |
| 584 | int status = 0; |
| 585 | |
| 586 | if (time_before(jiffies, req->rq_majortimeo)) { |
| 587 | if (to->to_exponential) |
| 588 | req->rq_timeout <<= 1; |
| 589 | else |
| 590 | req->rq_timeout += to->to_increment; |
| 591 | if (to->to_maxval && req->rq_timeout >= to->to_maxval) |
| 592 | req->rq_timeout = to->to_maxval; |
| 593 | req->rq_retries++; |
| 594 | } else { |
| 595 | req->rq_timeout = to->to_initval; |
| 596 | req->rq_retries = 0; |
| 597 | xprt_reset_majortimeo(req); |
| 598 | /* Reset the RTT counters == "slow start" */ |
| 599 | spin_lock_bh(&xprt->transport_lock); |
| 600 | rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval); |
| 601 | spin_unlock_bh(&xprt->transport_lock); |
| 602 | status = -ETIMEDOUT; |
| 603 | } |
| 604 | |
| 605 | if (req->rq_timeout == 0) { |
| 606 | printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n"); |
| 607 | req->rq_timeout = 5 * HZ; |
| 608 | } |
| 609 | return status; |
| 610 | } |
| 611 | |
| 612 | static void xprt_autoclose(struct work_struct *work) |
| 613 | { |
| 614 | struct rpc_xprt *xprt = |
| 615 | container_of(work, struct rpc_xprt, task_cleanup); |
| 616 | |
| 617 | clear_bit(XPRT_CLOSE_WAIT, &xprt->state); |
| 618 | xprt->ops->close(xprt); |
| 619 | xprt_release_write(xprt, NULL); |
| 620 | wake_up_bit(&xprt->state, XPRT_LOCKED); |
| 621 | } |
| 622 | |
| 623 | /** |
| 624 | * xprt_disconnect_done - mark a transport as disconnected |
| 625 | * @xprt: transport to flag for disconnect |
| 626 | * |
| 627 | */ |
| 628 | void xprt_disconnect_done(struct rpc_xprt *xprt) |
| 629 | { |
| 630 | dprintk("RPC: disconnected transport %p\n", xprt); |
| 631 | spin_lock_bh(&xprt->transport_lock); |
| 632 | xprt_clear_connected(xprt); |
| 633 | xprt_wake_pending_tasks(xprt, -EAGAIN); |
| 634 | spin_unlock_bh(&xprt->transport_lock); |
| 635 | } |
| 636 | EXPORT_SYMBOL_GPL(xprt_disconnect_done); |
| 637 | |
| 638 | /** |
| 639 | * xprt_force_disconnect - force a transport to disconnect |
| 640 | * @xprt: transport to disconnect |
| 641 | * |
| 642 | */ |
| 643 | void xprt_force_disconnect(struct rpc_xprt *xprt) |
| 644 | { |
| 645 | /* Don't race with the test_bit() in xprt_clear_locked() */ |
| 646 | spin_lock_bh(&xprt->transport_lock); |
| 647 | set_bit(XPRT_CLOSE_WAIT, &xprt->state); |
| 648 | /* Try to schedule an autoclose RPC call */ |
| 649 | if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0) |
| 650 | queue_work(xprtiod_workqueue, &xprt->task_cleanup); |
| 651 | xprt_wake_pending_tasks(xprt, -EAGAIN); |
| 652 | spin_unlock_bh(&xprt->transport_lock); |
| 653 | } |
| 654 | EXPORT_SYMBOL_GPL(xprt_force_disconnect); |
| 655 | |
| 656 | /** |
| 657 | * xprt_conditional_disconnect - force a transport to disconnect |
| 658 | * @xprt: transport to disconnect |
| 659 | * @cookie: 'connection cookie' |
| 660 | * |
| 661 | * This attempts to break the connection if and only if 'cookie' matches |
| 662 | * the current transport 'connection cookie'. It ensures that we don't |
| 663 | * try to break the connection more than once when we need to retransmit |
| 664 | * a batch of RPC requests. |
| 665 | * |
| 666 | */ |
| 667 | void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie) |
| 668 | { |
| 669 | /* Don't race with the test_bit() in xprt_clear_locked() */ |
| 670 | spin_lock_bh(&xprt->transport_lock); |
| 671 | if (cookie != xprt->connect_cookie) |
| 672 | goto out; |
| 673 | if (test_bit(XPRT_CLOSING, &xprt->state)) |
| 674 | goto out; |
| 675 | set_bit(XPRT_CLOSE_WAIT, &xprt->state); |
| 676 | /* Try to schedule an autoclose RPC call */ |
| 677 | if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0) |
| 678 | queue_work(xprtiod_workqueue, &xprt->task_cleanup); |
| 679 | xprt_wake_pending_tasks(xprt, -EAGAIN); |
| 680 | out: |
| 681 | spin_unlock_bh(&xprt->transport_lock); |
| 682 | } |
| 683 | |
| 684 | static bool |
| 685 | xprt_has_timer(const struct rpc_xprt *xprt) |
| 686 | { |
| 687 | return xprt->idle_timeout != 0; |
| 688 | } |
| 689 | |
| 690 | static void |
| 691 | xprt_schedule_autodisconnect(struct rpc_xprt *xprt) |
| 692 | __must_hold(&xprt->transport_lock) |
| 693 | { |
| 694 | if (list_empty(&xprt->recv) && xprt_has_timer(xprt)) |
| 695 | mod_timer(&xprt->timer, xprt->last_used + xprt->idle_timeout); |
| 696 | } |
| 697 | |
| 698 | static void |
| 699 | xprt_init_autodisconnect(unsigned long data) |
| 700 | { |
| 701 | struct rpc_xprt *xprt = (struct rpc_xprt *)data; |
| 702 | |
| 703 | spin_lock(&xprt->transport_lock); |
| 704 | if (!list_empty(&xprt->recv)) |
| 705 | goto out_abort; |
| 706 | /* Reset xprt->last_used to avoid connect/autodisconnect cycling */ |
| 707 | xprt->last_used = jiffies; |
| 708 | if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) |
| 709 | goto out_abort; |
| 710 | spin_unlock(&xprt->transport_lock); |
| 711 | queue_work(xprtiod_workqueue, &xprt->task_cleanup); |
| 712 | return; |
| 713 | out_abort: |
| 714 | spin_unlock(&xprt->transport_lock); |
| 715 | } |
| 716 | |
| 717 | bool xprt_lock_connect(struct rpc_xprt *xprt, |
| 718 | struct rpc_task *task, |
| 719 | void *cookie) |
| 720 | { |
| 721 | bool ret = false; |
| 722 | |
| 723 | spin_lock_bh(&xprt->transport_lock); |
| 724 | if (!test_bit(XPRT_LOCKED, &xprt->state)) |
| 725 | goto out; |
| 726 | if (xprt->snd_task != task) |
| 727 | goto out; |
| 728 | xprt_task_clear_bytes_sent(task); |
| 729 | xprt->snd_task = cookie; |
| 730 | ret = true; |
| 731 | out: |
| 732 | spin_unlock_bh(&xprt->transport_lock); |
| 733 | return ret; |
| 734 | } |
| 735 | |
| 736 | void xprt_unlock_connect(struct rpc_xprt *xprt, void *cookie) |
| 737 | { |
| 738 | spin_lock_bh(&xprt->transport_lock); |
| 739 | if (xprt->snd_task != cookie) |
| 740 | goto out; |
| 741 | if (!test_bit(XPRT_LOCKED, &xprt->state)) |
| 742 | goto out; |
| 743 | xprt->snd_task =NULL; |
| 744 | xprt->ops->release_xprt(xprt, NULL); |
| 745 | xprt_schedule_autodisconnect(xprt); |
| 746 | out: |
| 747 | spin_unlock_bh(&xprt->transport_lock); |
| 748 | wake_up_bit(&xprt->state, XPRT_LOCKED); |
| 749 | } |
| 750 | |
| 751 | /** |
| 752 | * xprt_connect - schedule a transport connect operation |
| 753 | * @task: RPC task that is requesting the connect |
| 754 | * |
| 755 | */ |
| 756 | void xprt_connect(struct rpc_task *task) |
| 757 | { |
| 758 | struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt; |
| 759 | |
| 760 | dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid, |
| 761 | xprt, (xprt_connected(xprt) ? "is" : "is not")); |
| 762 | |
| 763 | if (!xprt_bound(xprt)) { |
| 764 | task->tk_status = -EAGAIN; |
| 765 | return; |
| 766 | } |
| 767 | if (!xprt_lock_write(xprt, task)) |
| 768 | return; |
| 769 | |
| 770 | if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state)) |
| 771 | xprt->ops->close(xprt); |
| 772 | |
| 773 | if (!xprt_connected(xprt)) { |
| 774 | task->tk_rqstp->rq_bytes_sent = 0; |
| 775 | task->tk_timeout = task->tk_rqstp->rq_timeout; |
| 776 | task->tk_rqstp->rq_connect_cookie = xprt->connect_cookie; |
| 777 | rpc_sleep_on(&xprt->pending, task, xprt_connect_status); |
| 778 | |
| 779 | if (test_bit(XPRT_CLOSING, &xprt->state)) |
| 780 | return; |
| 781 | if (xprt_test_and_set_connecting(xprt)) |
| 782 | return; |
| 783 | /* Race breaker */ |
| 784 | if (!xprt_connected(xprt)) { |
| 785 | xprt->stat.connect_start = jiffies; |
| 786 | xprt->ops->connect(xprt, task); |
| 787 | } else { |
| 788 | xprt_clear_connecting(xprt); |
| 789 | task->tk_status = 0; |
| 790 | rpc_wake_up_queued_task(&xprt->pending, task); |
| 791 | } |
| 792 | } |
| 793 | xprt_release_write(xprt, task); |
| 794 | } |
| 795 | |
| 796 | static void xprt_connect_status(struct rpc_task *task) |
| 797 | { |
| 798 | struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt; |
| 799 | |
| 800 | if (task->tk_status == 0) { |
| 801 | xprt->stat.connect_count++; |
| 802 | xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start; |
| 803 | dprintk("RPC: %5u xprt_connect_status: connection established\n", |
| 804 | task->tk_pid); |
| 805 | return; |
| 806 | } |
| 807 | |
| 808 | switch (task->tk_status) { |
| 809 | case -ECONNREFUSED: |
| 810 | case -ECONNRESET: |
| 811 | case -ECONNABORTED: |
| 812 | case -ENETUNREACH: |
| 813 | case -EHOSTUNREACH: |
| 814 | case -EPIPE: |
| 815 | case -EAGAIN: |
| 816 | dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid); |
| 817 | break; |
| 818 | case -ETIMEDOUT: |
| 819 | dprintk("RPC: %5u xprt_connect_status: connect attempt timed " |
| 820 | "out\n", task->tk_pid); |
| 821 | break; |
| 822 | default: |
| 823 | dprintk("RPC: %5u xprt_connect_status: error %d connecting to " |
| 824 | "server %s\n", task->tk_pid, -task->tk_status, |
| 825 | xprt->servername); |
| 826 | task->tk_status = -EIO; |
| 827 | } |
| 828 | } |
| 829 | |
| 830 | /** |
| 831 | * xprt_lookup_rqst - find an RPC request corresponding to an XID |
| 832 | * @xprt: transport on which the original request was transmitted |
| 833 | * @xid: RPC XID of incoming reply |
| 834 | * |
| 835 | */ |
| 836 | struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid) |
| 837 | { |
| 838 | struct rpc_rqst *entry; |
| 839 | |
| 840 | list_for_each_entry(entry, &xprt->recv, rq_list) |
| 841 | if (entry->rq_xid == xid) { |
| 842 | trace_xprt_lookup_rqst(xprt, xid, 0); |
| 843 | return entry; |
| 844 | } |
| 845 | |
| 846 | dprintk("RPC: xprt_lookup_rqst did not find xid %08x\n", |
| 847 | ntohl(xid)); |
| 848 | trace_xprt_lookup_rqst(xprt, xid, -ENOENT); |
| 849 | xprt->stat.bad_xids++; |
| 850 | return NULL; |
| 851 | } |
| 852 | EXPORT_SYMBOL_GPL(xprt_lookup_rqst); |
| 853 | |
| 854 | /** |
| 855 | * xprt_pin_rqst - Pin a request on the transport receive list |
| 856 | * @req: Request to pin |
| 857 | * |
| 858 | * Caller must ensure this is atomic with the call to xprt_lookup_rqst() |
| 859 | * so should be holding the xprt transport lock. |
| 860 | */ |
| 861 | void xprt_pin_rqst(struct rpc_rqst *req) |
| 862 | { |
| 863 | set_bit(RPC_TASK_MSG_RECV, &req->rq_task->tk_runstate); |
| 864 | } |
| 865 | EXPORT_SYMBOL_GPL(xprt_pin_rqst); |
| 866 | |
| 867 | /** |
| 868 | * xprt_unpin_rqst - Unpin a request on the transport receive list |
| 869 | * @req: Request to pin |
| 870 | * |
| 871 | * Caller should be holding the xprt transport lock. |
| 872 | */ |
| 873 | void xprt_unpin_rqst(struct rpc_rqst *req) |
| 874 | { |
| 875 | struct rpc_task *task = req->rq_task; |
| 876 | |
| 877 | clear_bit(RPC_TASK_MSG_RECV, &task->tk_runstate); |
| 878 | if (test_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate)) |
| 879 | wake_up_bit(&task->tk_runstate, RPC_TASK_MSG_RECV); |
| 880 | } |
| 881 | EXPORT_SYMBOL_GPL(xprt_unpin_rqst); |
| 882 | |
| 883 | static void xprt_wait_on_pinned_rqst(struct rpc_rqst *req) |
| 884 | __must_hold(&req->rq_xprt->recv_lock) |
| 885 | { |
| 886 | struct rpc_task *task = req->rq_task; |
| 887 | |
| 888 | if (task && test_bit(RPC_TASK_MSG_RECV, &task->tk_runstate)) { |
| 889 | spin_unlock(&req->rq_xprt->recv_lock); |
| 890 | set_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate); |
| 891 | wait_on_bit(&task->tk_runstate, RPC_TASK_MSG_RECV, |
| 892 | TASK_UNINTERRUPTIBLE); |
| 893 | clear_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate); |
| 894 | spin_lock(&req->rq_xprt->recv_lock); |
| 895 | } |
| 896 | } |
| 897 | |
| 898 | static void xprt_update_rtt(struct rpc_task *task) |
| 899 | { |
| 900 | struct rpc_rqst *req = task->tk_rqstp; |
| 901 | struct rpc_rtt *rtt = task->tk_client->cl_rtt; |
| 902 | unsigned int timer = task->tk_msg.rpc_proc->p_timer; |
| 903 | long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt)); |
| 904 | |
| 905 | if (timer) { |
| 906 | if (req->rq_ntrans == 1) |
| 907 | rpc_update_rtt(rtt, timer, m); |
| 908 | rpc_set_timeo(rtt, timer, req->rq_ntrans - 1); |
| 909 | } |
| 910 | } |
| 911 | |
| 912 | /** |
| 913 | * xprt_complete_rqst - called when reply processing is complete |
| 914 | * @task: RPC request that recently completed |
| 915 | * @copied: actual number of bytes received from the transport |
| 916 | * |
| 917 | * Caller holds transport lock. |
| 918 | */ |
| 919 | void xprt_complete_rqst(struct rpc_task *task, int copied) |
| 920 | { |
| 921 | struct rpc_rqst *req = task->tk_rqstp; |
| 922 | struct rpc_xprt *xprt = req->rq_xprt; |
| 923 | |
| 924 | dprintk("RPC: %5u xid %08x complete (%d bytes received)\n", |
| 925 | task->tk_pid, ntohl(req->rq_xid), copied); |
| 926 | trace_xprt_complete_rqst(xprt, req->rq_xid, copied); |
| 927 | |
| 928 | xprt->stat.recvs++; |
| 929 | req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime); |
| 930 | if (xprt->ops->timer != NULL) |
| 931 | xprt_update_rtt(task); |
| 932 | |
| 933 | list_del_init(&req->rq_list); |
| 934 | req->rq_private_buf.len = copied; |
| 935 | /* Ensure all writes are done before we update */ |
| 936 | /* req->rq_reply_bytes_recvd */ |
| 937 | smp_wmb(); |
| 938 | req->rq_reply_bytes_recvd = copied; |
| 939 | rpc_wake_up_queued_task(&xprt->pending, task); |
| 940 | } |
| 941 | EXPORT_SYMBOL_GPL(xprt_complete_rqst); |
| 942 | |
| 943 | static void xprt_timer(struct rpc_task *task) |
| 944 | { |
| 945 | struct rpc_rqst *req = task->tk_rqstp; |
| 946 | struct rpc_xprt *xprt = req->rq_xprt; |
| 947 | |
| 948 | if (task->tk_status != -ETIMEDOUT) |
| 949 | return; |
| 950 | dprintk("RPC: %5u xprt_timer\n", task->tk_pid); |
| 951 | |
| 952 | if (!req->rq_reply_bytes_recvd) { |
| 953 | if (xprt->ops->timer) |
| 954 | xprt->ops->timer(xprt, task); |
| 955 | } else |
| 956 | task->tk_status = 0; |
| 957 | } |
| 958 | |
| 959 | /** |
| 960 | * xprt_prepare_transmit - reserve the transport before sending a request |
| 961 | * @task: RPC task about to send a request |
| 962 | * |
| 963 | */ |
| 964 | bool xprt_prepare_transmit(struct rpc_task *task) |
| 965 | { |
| 966 | struct rpc_rqst *req = task->tk_rqstp; |
| 967 | struct rpc_xprt *xprt = req->rq_xprt; |
| 968 | bool ret = false; |
| 969 | |
| 970 | dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid); |
| 971 | |
| 972 | spin_lock_bh(&xprt->transport_lock); |
| 973 | if (!req->rq_bytes_sent) { |
| 974 | if (req->rq_reply_bytes_recvd) { |
| 975 | task->tk_status = req->rq_reply_bytes_recvd; |
| 976 | goto out_unlock; |
| 977 | } |
| 978 | if ((task->tk_flags & RPC_TASK_NO_RETRANS_TIMEOUT) |
| 979 | && xprt_connected(xprt) |
| 980 | && req->rq_connect_cookie == xprt->connect_cookie) { |
| 981 | xprt->ops->set_retrans_timeout(task); |
| 982 | rpc_sleep_on(&xprt->pending, task, xprt_timer); |
| 983 | goto out_unlock; |
| 984 | } |
| 985 | } |
| 986 | if (!xprt->ops->reserve_xprt(xprt, task)) { |
| 987 | task->tk_status = -EAGAIN; |
| 988 | goto out_unlock; |
| 989 | } |
| 990 | ret = true; |
| 991 | out_unlock: |
| 992 | spin_unlock_bh(&xprt->transport_lock); |
| 993 | return ret; |
| 994 | } |
| 995 | |
| 996 | void xprt_end_transmit(struct rpc_task *task) |
| 997 | { |
| 998 | xprt_release_write(task->tk_rqstp->rq_xprt, task); |
| 999 | } |
| 1000 | |
| 1001 | /** |
| 1002 | * xprt_transmit - send an RPC request on a transport |
| 1003 | * @task: controlling RPC task |
| 1004 | * |
| 1005 | * We have to copy the iovec because sendmsg fiddles with its contents. |
| 1006 | */ |
| 1007 | void xprt_transmit(struct rpc_task *task) |
| 1008 | { |
| 1009 | struct rpc_rqst *req = task->tk_rqstp; |
| 1010 | struct rpc_xprt *xprt = req->rq_xprt; |
| 1011 | unsigned int connect_cookie; |
| 1012 | int status, numreqs; |
| 1013 | |
| 1014 | dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen); |
| 1015 | |
| 1016 | if (!req->rq_reply_bytes_recvd) { |
| 1017 | if (list_empty(&req->rq_list) && rpc_reply_expected(task)) { |
| 1018 | /* |
| 1019 | * Add to the list only if we're expecting a reply |
| 1020 | */ |
| 1021 | /* Update the softirq receive buffer */ |
| 1022 | memcpy(&req->rq_private_buf, &req->rq_rcv_buf, |
| 1023 | sizeof(req->rq_private_buf)); |
| 1024 | /* Add request to the receive list */ |
| 1025 | spin_lock(&xprt->recv_lock); |
| 1026 | list_add_tail(&req->rq_list, &xprt->recv); |
| 1027 | spin_unlock(&xprt->recv_lock); |
| 1028 | xprt_reset_majortimeo(req); |
| 1029 | /* Turn off autodisconnect */ |
| 1030 | del_singleshot_timer_sync(&xprt->timer); |
| 1031 | } |
| 1032 | } else if (!req->rq_bytes_sent) |
| 1033 | return; |
| 1034 | |
| 1035 | connect_cookie = xprt->connect_cookie; |
| 1036 | req->rq_xtime = ktime_get(); |
| 1037 | status = xprt->ops->send_request(task); |
| 1038 | trace_xprt_transmit(xprt, req->rq_xid, status); |
| 1039 | if (status != 0) { |
| 1040 | task->tk_status = status; |
| 1041 | return; |
| 1042 | } |
| 1043 | xprt_inject_disconnect(xprt); |
| 1044 | |
| 1045 | dprintk("RPC: %5u xmit complete\n", task->tk_pid); |
| 1046 | task->tk_flags |= RPC_TASK_SENT; |
| 1047 | spin_lock_bh(&xprt->transport_lock); |
| 1048 | |
| 1049 | xprt->ops->set_retrans_timeout(task); |
| 1050 | |
| 1051 | numreqs = atomic_read(&xprt->num_reqs); |
| 1052 | if (numreqs > xprt->stat.max_slots) |
| 1053 | xprt->stat.max_slots = numreqs; |
| 1054 | xprt->stat.sends++; |
| 1055 | xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs; |
| 1056 | xprt->stat.bklog_u += xprt->backlog.qlen; |
| 1057 | xprt->stat.sending_u += xprt->sending.qlen; |
| 1058 | xprt->stat.pending_u += xprt->pending.qlen; |
| 1059 | spin_unlock_bh(&xprt->transport_lock); |
| 1060 | |
| 1061 | req->rq_connect_cookie = connect_cookie; |
| 1062 | if (rpc_reply_expected(task) && !READ_ONCE(req->rq_reply_bytes_recvd)) { |
| 1063 | /* |
| 1064 | * Sleep on the pending queue if we're expecting a reply. |
| 1065 | * The spinlock ensures atomicity between the test of |
| 1066 | * req->rq_reply_bytes_recvd, and the call to rpc_sleep_on(). |
| 1067 | */ |
| 1068 | spin_lock(&xprt->recv_lock); |
| 1069 | if (!req->rq_reply_bytes_recvd) { |
| 1070 | rpc_sleep_on(&xprt->pending, task, xprt_timer); |
| 1071 | /* |
| 1072 | * Send an extra queue wakeup call if the |
| 1073 | * connection was dropped in case the call to |
| 1074 | * rpc_sleep_on() raced. |
| 1075 | */ |
| 1076 | if (!xprt_connected(xprt)) |
| 1077 | xprt_wake_pending_tasks(xprt, -ENOTCONN); |
| 1078 | } |
| 1079 | spin_unlock(&xprt->recv_lock); |
| 1080 | } |
| 1081 | } |
| 1082 | |
| 1083 | static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task) |
| 1084 | { |
| 1085 | set_bit(XPRT_CONGESTED, &xprt->state); |
| 1086 | rpc_sleep_on(&xprt->backlog, task, NULL); |
| 1087 | } |
| 1088 | |
| 1089 | static void xprt_wake_up_backlog(struct rpc_xprt *xprt) |
| 1090 | { |
| 1091 | if (rpc_wake_up_next(&xprt->backlog) == NULL) |
| 1092 | clear_bit(XPRT_CONGESTED, &xprt->state); |
| 1093 | } |
| 1094 | |
| 1095 | static bool xprt_throttle_congested(struct rpc_xprt *xprt, struct rpc_task *task) |
| 1096 | { |
| 1097 | bool ret = false; |
| 1098 | |
| 1099 | if (!test_bit(XPRT_CONGESTED, &xprt->state)) |
| 1100 | goto out; |
| 1101 | spin_lock(&xprt->reserve_lock); |
| 1102 | if (test_bit(XPRT_CONGESTED, &xprt->state)) { |
| 1103 | rpc_sleep_on(&xprt->backlog, task, NULL); |
| 1104 | ret = true; |
| 1105 | } |
| 1106 | spin_unlock(&xprt->reserve_lock); |
| 1107 | out: |
| 1108 | return ret; |
| 1109 | } |
| 1110 | |
| 1111 | static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt) |
| 1112 | { |
| 1113 | struct rpc_rqst *req = ERR_PTR(-EAGAIN); |
| 1114 | |
| 1115 | if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs)) |
| 1116 | goto out; |
| 1117 | spin_unlock(&xprt->reserve_lock); |
| 1118 | req = kzalloc(sizeof(struct rpc_rqst), GFP_NOFS); |
| 1119 | spin_lock(&xprt->reserve_lock); |
| 1120 | if (req != NULL) |
| 1121 | goto out; |
| 1122 | atomic_dec(&xprt->num_reqs); |
| 1123 | req = ERR_PTR(-ENOMEM); |
| 1124 | out: |
| 1125 | return req; |
| 1126 | } |
| 1127 | |
| 1128 | static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req) |
| 1129 | { |
| 1130 | if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) { |
| 1131 | kfree(req); |
| 1132 | return true; |
| 1133 | } |
| 1134 | return false; |
| 1135 | } |
| 1136 | |
| 1137 | void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task) |
| 1138 | { |
| 1139 | struct rpc_rqst *req; |
| 1140 | |
| 1141 | spin_lock(&xprt->reserve_lock); |
| 1142 | if (!list_empty(&xprt->free)) { |
| 1143 | req = list_entry(xprt->free.next, struct rpc_rqst, rq_list); |
| 1144 | list_del(&req->rq_list); |
| 1145 | goto out_init_req; |
| 1146 | } |
| 1147 | req = xprt_dynamic_alloc_slot(xprt); |
| 1148 | if (!IS_ERR(req)) |
| 1149 | goto out_init_req; |
| 1150 | switch (PTR_ERR(req)) { |
| 1151 | case -ENOMEM: |
| 1152 | dprintk("RPC: dynamic allocation of request slot " |
| 1153 | "failed! Retrying\n"); |
| 1154 | task->tk_status = -ENOMEM; |
| 1155 | break; |
| 1156 | case -EAGAIN: |
| 1157 | xprt_add_backlog(xprt, task); |
| 1158 | dprintk("RPC: waiting for request slot\n"); |
| 1159 | default: |
| 1160 | task->tk_status = -EAGAIN; |
| 1161 | } |
| 1162 | spin_unlock(&xprt->reserve_lock); |
| 1163 | return; |
| 1164 | out_init_req: |
| 1165 | task->tk_status = 0; |
| 1166 | task->tk_rqstp = req; |
| 1167 | xprt_request_init(task, xprt); |
| 1168 | spin_unlock(&xprt->reserve_lock); |
| 1169 | } |
| 1170 | EXPORT_SYMBOL_GPL(xprt_alloc_slot); |
| 1171 | |
| 1172 | void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task) |
| 1173 | { |
| 1174 | /* Note: grabbing the xprt_lock_write() ensures that we throttle |
| 1175 | * new slot allocation if the transport is congested (i.e. when |
| 1176 | * reconnecting a stream transport or when out of socket write |
| 1177 | * buffer space). |
| 1178 | */ |
| 1179 | if (xprt_lock_write(xprt, task)) { |
| 1180 | xprt_alloc_slot(xprt, task); |
| 1181 | xprt_release_write(xprt, task); |
| 1182 | } |
| 1183 | } |
| 1184 | EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot); |
| 1185 | |
| 1186 | static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req) |
| 1187 | { |
| 1188 | spin_lock(&xprt->reserve_lock); |
| 1189 | if (!xprt_dynamic_free_slot(xprt, req)) { |
| 1190 | memset(req, 0, sizeof(*req)); /* mark unused */ |
| 1191 | list_add(&req->rq_list, &xprt->free); |
| 1192 | } |
| 1193 | xprt_wake_up_backlog(xprt); |
| 1194 | spin_unlock(&xprt->reserve_lock); |
| 1195 | } |
| 1196 | |
| 1197 | static void xprt_free_all_slots(struct rpc_xprt *xprt) |
| 1198 | { |
| 1199 | struct rpc_rqst *req; |
| 1200 | while (!list_empty(&xprt->free)) { |
| 1201 | req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list); |
| 1202 | list_del(&req->rq_list); |
| 1203 | kfree(req); |
| 1204 | } |
| 1205 | } |
| 1206 | |
| 1207 | struct rpc_xprt *xprt_alloc(struct net *net, size_t size, |
| 1208 | unsigned int num_prealloc, |
| 1209 | unsigned int max_alloc) |
| 1210 | { |
| 1211 | struct rpc_xprt *xprt; |
| 1212 | struct rpc_rqst *req; |
| 1213 | int i; |
| 1214 | |
| 1215 | xprt = kzalloc(size, GFP_KERNEL); |
| 1216 | if (xprt == NULL) |
| 1217 | goto out; |
| 1218 | |
| 1219 | xprt_init(xprt, net); |
| 1220 | |
| 1221 | for (i = 0; i < num_prealloc; i++) { |
| 1222 | req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL); |
| 1223 | if (!req) |
| 1224 | goto out_free; |
| 1225 | list_add(&req->rq_list, &xprt->free); |
| 1226 | } |
| 1227 | if (max_alloc > num_prealloc) |
| 1228 | xprt->max_reqs = max_alloc; |
| 1229 | else |
| 1230 | xprt->max_reqs = num_prealloc; |
| 1231 | xprt->min_reqs = num_prealloc; |
| 1232 | atomic_set(&xprt->num_reqs, num_prealloc); |
| 1233 | |
| 1234 | return xprt; |
| 1235 | |
| 1236 | out_free: |
| 1237 | xprt_free(xprt); |
| 1238 | out: |
| 1239 | return NULL; |
| 1240 | } |
| 1241 | EXPORT_SYMBOL_GPL(xprt_alloc); |
| 1242 | |
| 1243 | void xprt_free(struct rpc_xprt *xprt) |
| 1244 | { |
| 1245 | put_net(xprt->xprt_net); |
| 1246 | xprt_free_all_slots(xprt); |
| 1247 | kfree_rcu(xprt, rcu); |
| 1248 | } |
| 1249 | EXPORT_SYMBOL_GPL(xprt_free); |
| 1250 | |
| 1251 | /** |
| 1252 | * xprt_reserve - allocate an RPC request slot |
| 1253 | * @task: RPC task requesting a slot allocation |
| 1254 | * |
| 1255 | * If the transport is marked as being congested, or if no more |
| 1256 | * slots are available, place the task on the transport's |
| 1257 | * backlog queue. |
| 1258 | */ |
| 1259 | void xprt_reserve(struct rpc_task *task) |
| 1260 | { |
| 1261 | struct rpc_xprt *xprt = task->tk_xprt; |
| 1262 | |
| 1263 | task->tk_status = 0; |
| 1264 | if (task->tk_rqstp != NULL) |
| 1265 | return; |
| 1266 | |
| 1267 | task->tk_timeout = 0; |
| 1268 | task->tk_status = -EAGAIN; |
| 1269 | if (!xprt_throttle_congested(xprt, task)) |
| 1270 | xprt->ops->alloc_slot(xprt, task); |
| 1271 | } |
| 1272 | |
| 1273 | /** |
| 1274 | * xprt_retry_reserve - allocate an RPC request slot |
| 1275 | * @task: RPC task requesting a slot allocation |
| 1276 | * |
| 1277 | * If no more slots are available, place the task on the transport's |
| 1278 | * backlog queue. |
| 1279 | * Note that the only difference with xprt_reserve is that we now |
| 1280 | * ignore the value of the XPRT_CONGESTED flag. |
| 1281 | */ |
| 1282 | void xprt_retry_reserve(struct rpc_task *task) |
| 1283 | { |
| 1284 | struct rpc_xprt *xprt = task->tk_xprt; |
| 1285 | |
| 1286 | task->tk_status = 0; |
| 1287 | if (task->tk_rqstp != NULL) |
| 1288 | return; |
| 1289 | |
| 1290 | task->tk_timeout = 0; |
| 1291 | task->tk_status = -EAGAIN; |
| 1292 | xprt->ops->alloc_slot(xprt, task); |
| 1293 | } |
| 1294 | |
| 1295 | static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt) |
| 1296 | { |
| 1297 | return (__force __be32)xprt->xid++; |
| 1298 | } |
| 1299 | |
| 1300 | static inline void xprt_init_xid(struct rpc_xprt *xprt) |
| 1301 | { |
| 1302 | xprt->xid = prandom_u32(); |
| 1303 | } |
| 1304 | |
| 1305 | static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt) |
| 1306 | { |
| 1307 | struct rpc_rqst *req = task->tk_rqstp; |
| 1308 | |
| 1309 | INIT_LIST_HEAD(&req->rq_list); |
| 1310 | req->rq_timeout = task->tk_client->cl_timeout->to_initval; |
| 1311 | req->rq_task = task; |
| 1312 | req->rq_xprt = xprt; |
| 1313 | req->rq_buffer = NULL; |
| 1314 | req->rq_xid = xprt_alloc_xid(xprt); |
| 1315 | req->rq_connect_cookie = xprt->connect_cookie - 1; |
| 1316 | req->rq_bytes_sent = 0; |
| 1317 | req->rq_snd_buf.len = 0; |
| 1318 | req->rq_snd_buf.buflen = 0; |
| 1319 | req->rq_rcv_buf.len = 0; |
| 1320 | req->rq_rcv_buf.buflen = 0; |
| 1321 | req->rq_release_snd_buf = NULL; |
| 1322 | xprt_reset_majortimeo(req); |
| 1323 | dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid, |
| 1324 | req, ntohl(req->rq_xid)); |
| 1325 | } |
| 1326 | |
| 1327 | /** |
| 1328 | * xprt_release - release an RPC request slot |
| 1329 | * @task: task which is finished with the slot |
| 1330 | * |
| 1331 | */ |
| 1332 | void xprt_release(struct rpc_task *task) |
| 1333 | { |
| 1334 | struct rpc_xprt *xprt; |
| 1335 | struct rpc_rqst *req = task->tk_rqstp; |
| 1336 | |
| 1337 | if (req == NULL) { |
| 1338 | if (task->tk_client) { |
| 1339 | xprt = task->tk_xprt; |
| 1340 | if (xprt->snd_task == task) |
| 1341 | xprt_release_write(xprt, task); |
| 1342 | } |
| 1343 | return; |
| 1344 | } |
| 1345 | |
| 1346 | xprt = req->rq_xprt; |
| 1347 | if (task->tk_ops->rpc_count_stats != NULL) |
| 1348 | task->tk_ops->rpc_count_stats(task, task->tk_calldata); |
| 1349 | else if (task->tk_client) |
| 1350 | rpc_count_iostats(task, task->tk_client->cl_metrics); |
| 1351 | spin_lock(&xprt->recv_lock); |
| 1352 | if (!list_empty(&req->rq_list)) { |
| 1353 | list_del_init(&req->rq_list); |
| 1354 | xprt_wait_on_pinned_rqst(req); |
| 1355 | } |
| 1356 | spin_unlock(&xprt->recv_lock); |
| 1357 | spin_lock_bh(&xprt->transport_lock); |
| 1358 | xprt->ops->release_xprt(xprt, task); |
| 1359 | if (xprt->ops->release_request) |
| 1360 | xprt->ops->release_request(task); |
| 1361 | xprt->last_used = jiffies; |
| 1362 | xprt_schedule_autodisconnect(xprt); |
| 1363 | spin_unlock_bh(&xprt->transport_lock); |
| 1364 | if (req->rq_buffer) |
| 1365 | xprt->ops->buf_free(task); |
| 1366 | xprt_inject_disconnect(xprt); |
| 1367 | if (req->rq_cred != NULL) |
| 1368 | put_rpccred(req->rq_cred); |
| 1369 | task->tk_rqstp = NULL; |
| 1370 | if (req->rq_release_snd_buf) |
| 1371 | req->rq_release_snd_buf(req); |
| 1372 | |
| 1373 | dprintk("RPC: %5u release request %p\n", task->tk_pid, req); |
| 1374 | if (likely(!bc_prealloc(req))) |
| 1375 | xprt_free_slot(xprt, req); |
| 1376 | else |
| 1377 | xprt_free_bc_request(req); |
| 1378 | } |
| 1379 | |
| 1380 | static void xprt_init(struct rpc_xprt *xprt, struct net *net) |
| 1381 | { |
| 1382 | kref_init(&xprt->kref); |
| 1383 | |
| 1384 | spin_lock_init(&xprt->transport_lock); |
| 1385 | spin_lock_init(&xprt->reserve_lock); |
| 1386 | spin_lock_init(&xprt->recv_lock); |
| 1387 | |
| 1388 | INIT_LIST_HEAD(&xprt->free); |
| 1389 | INIT_LIST_HEAD(&xprt->recv); |
| 1390 | #if defined(CONFIG_SUNRPC_BACKCHANNEL) |
| 1391 | spin_lock_init(&xprt->bc_pa_lock); |
| 1392 | INIT_LIST_HEAD(&xprt->bc_pa_list); |
| 1393 | #endif /* CONFIG_SUNRPC_BACKCHANNEL */ |
| 1394 | INIT_LIST_HEAD(&xprt->xprt_switch); |
| 1395 | |
| 1396 | xprt->last_used = jiffies; |
| 1397 | xprt->cwnd = RPC_INITCWND; |
| 1398 | xprt->bind_index = 0; |
| 1399 | |
| 1400 | rpc_init_wait_queue(&xprt->binding, "xprt_binding"); |
| 1401 | rpc_init_wait_queue(&xprt->pending, "xprt_pending"); |
| 1402 | rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending"); |
| 1403 | rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog"); |
| 1404 | |
| 1405 | xprt_init_xid(xprt); |
| 1406 | |
| 1407 | xprt->xprt_net = get_net(net); |
| 1408 | } |
| 1409 | |
| 1410 | /** |
| 1411 | * xprt_create_transport - create an RPC transport |
| 1412 | * @args: rpc transport creation arguments |
| 1413 | * |
| 1414 | */ |
| 1415 | struct rpc_xprt *xprt_create_transport(struct xprt_create *args) |
| 1416 | { |
| 1417 | struct rpc_xprt *xprt; |
| 1418 | struct xprt_class *t; |
| 1419 | |
| 1420 | spin_lock(&xprt_list_lock); |
| 1421 | list_for_each_entry(t, &xprt_list, list) { |
| 1422 | if (t->ident == args->ident) { |
| 1423 | spin_unlock(&xprt_list_lock); |
| 1424 | goto found; |
| 1425 | } |
| 1426 | } |
| 1427 | spin_unlock(&xprt_list_lock); |
| 1428 | dprintk("RPC: transport (%d) not supported\n", args->ident); |
| 1429 | return ERR_PTR(-EIO); |
| 1430 | |
| 1431 | found: |
| 1432 | xprt = t->setup(args); |
| 1433 | if (IS_ERR(xprt)) { |
| 1434 | dprintk("RPC: xprt_create_transport: failed, %ld\n", |
| 1435 | -PTR_ERR(xprt)); |
| 1436 | goto out; |
| 1437 | } |
| 1438 | if (args->flags & XPRT_CREATE_NO_IDLE_TIMEOUT) |
| 1439 | xprt->idle_timeout = 0; |
| 1440 | INIT_WORK(&xprt->task_cleanup, xprt_autoclose); |
| 1441 | if (xprt_has_timer(xprt)) |
| 1442 | setup_timer(&xprt->timer, xprt_init_autodisconnect, |
| 1443 | (unsigned long)xprt); |
| 1444 | else |
| 1445 | init_timer(&xprt->timer); |
| 1446 | |
| 1447 | if (strlen(args->servername) > RPC_MAXNETNAMELEN) { |
| 1448 | xprt_destroy(xprt); |
| 1449 | return ERR_PTR(-EINVAL); |
| 1450 | } |
| 1451 | xprt->servername = kstrdup(args->servername, GFP_KERNEL); |
| 1452 | if (xprt->servername == NULL) { |
| 1453 | xprt_destroy(xprt); |
| 1454 | return ERR_PTR(-ENOMEM); |
| 1455 | } |
| 1456 | |
| 1457 | rpc_xprt_debugfs_register(xprt); |
| 1458 | |
| 1459 | dprintk("RPC: created transport %p with %u slots\n", xprt, |
| 1460 | xprt->max_reqs); |
| 1461 | out: |
| 1462 | return xprt; |
| 1463 | } |
| 1464 | |
| 1465 | static void xprt_destroy_cb(struct work_struct *work) |
| 1466 | { |
| 1467 | struct rpc_xprt *xprt = |
| 1468 | container_of(work, struct rpc_xprt, task_cleanup); |
| 1469 | |
| 1470 | rpc_xprt_debugfs_unregister(xprt); |
| 1471 | rpc_destroy_wait_queue(&xprt->binding); |
| 1472 | rpc_destroy_wait_queue(&xprt->pending); |
| 1473 | rpc_destroy_wait_queue(&xprt->sending); |
| 1474 | rpc_destroy_wait_queue(&xprt->backlog); |
| 1475 | kfree(xprt->servername); |
| 1476 | /* |
| 1477 | * Tear down transport state and free the rpc_xprt |
| 1478 | */ |
| 1479 | xprt->ops->destroy(xprt); |
| 1480 | } |
| 1481 | |
| 1482 | /** |
| 1483 | * xprt_destroy - destroy an RPC transport, killing off all requests. |
| 1484 | * @xprt: transport to destroy |
| 1485 | * |
| 1486 | */ |
| 1487 | static void xprt_destroy(struct rpc_xprt *xprt) |
| 1488 | { |
| 1489 | dprintk("RPC: destroying transport %p\n", xprt); |
| 1490 | |
| 1491 | /* |
| 1492 | * Exclude transport connect/disconnect handlers and autoclose |
| 1493 | */ |
| 1494 | wait_on_bit_lock(&xprt->state, XPRT_LOCKED, TASK_UNINTERRUPTIBLE); |
| 1495 | |
| 1496 | del_timer_sync(&xprt->timer); |
| 1497 | |
| 1498 | /* |
| 1499 | * Destroy sockets etc from the system workqueue so they can |
| 1500 | * safely flush receive work running on rpciod. |
| 1501 | */ |
| 1502 | INIT_WORK(&xprt->task_cleanup, xprt_destroy_cb); |
| 1503 | schedule_work(&xprt->task_cleanup); |
| 1504 | } |
| 1505 | |
| 1506 | static void xprt_destroy_kref(struct kref *kref) |
| 1507 | { |
| 1508 | xprt_destroy(container_of(kref, struct rpc_xprt, kref)); |
| 1509 | } |
| 1510 | |
| 1511 | /** |
| 1512 | * xprt_get - return a reference to an RPC transport. |
| 1513 | * @xprt: pointer to the transport |
| 1514 | * |
| 1515 | */ |
| 1516 | struct rpc_xprt *xprt_get(struct rpc_xprt *xprt) |
| 1517 | { |
| 1518 | if (xprt != NULL && kref_get_unless_zero(&xprt->kref)) |
| 1519 | return xprt; |
| 1520 | return NULL; |
| 1521 | } |
| 1522 | EXPORT_SYMBOL_GPL(xprt_get); |
| 1523 | |
| 1524 | /** |
| 1525 | * xprt_put - release a reference to an RPC transport. |
| 1526 | * @xprt: pointer to the transport |
| 1527 | * |
| 1528 | */ |
| 1529 | void xprt_put(struct rpc_xprt *xprt) |
| 1530 | { |
| 1531 | if (xprt != NULL) |
| 1532 | kref_put(&xprt->kref, xprt_destroy_kref); |
| 1533 | } |
| 1534 | EXPORT_SYMBOL_GPL(xprt_put); |