Commit | Line | Data |
---|---|---|
f58851e6 TT |
1 | /* |
2 | * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved. | |
3 | * | |
4 | * This software is available to you under a choice of one of two | |
5 | * licenses. You may choose to be licensed under the terms of the GNU | |
6 | * General Public License (GPL) Version 2, available from the file | |
7 | * COPYING in the main directory of this source tree, or the BSD-type | |
8 | * license below: | |
9 | * | |
10 | * Redistribution and use in source and binary forms, with or without | |
11 | * modification, are permitted provided that the following conditions | |
12 | * are met: | |
13 | * | |
14 | * Redistributions of source code must retain the above copyright | |
15 | * notice, this list of conditions and the following disclaimer. | |
16 | * | |
17 | * Redistributions in binary form must reproduce the above | |
18 | * copyright notice, this list of conditions and the following | |
19 | * disclaimer in the documentation and/or other materials provided | |
20 | * with the distribution. | |
21 | * | |
22 | * Neither the name of the Network Appliance, Inc. nor the names of | |
23 | * its contributors may be used to endorse or promote products | |
24 | * derived from this software without specific prior written | |
25 | * permission. | |
26 | * | |
27 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
28 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
29 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR | |
30 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT | |
31 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, | |
32 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT | |
33 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, | |
34 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY | |
35 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT | |
36 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | |
37 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |
38 | */ | |
39 | ||
40 | /* | |
41 | * transport.c | |
42 | * | |
43 | * This file contains the top-level implementation of an RPC RDMA | |
44 | * transport. | |
45 | * | |
46 | * Naming convention: functions beginning with xprt_ are part of the | |
47 | * transport switch. All others are RPC RDMA internal. | |
48 | */ | |
49 | ||
50 | #include <linux/module.h> | |
51 | #include <linux/init.h> | |
52 | #include <linux/seq_file.h> | |
53 | ||
54 | #include "xprt_rdma.h" | |
55 | ||
56 | #ifdef RPC_DEBUG | |
57 | # define RPCDBG_FACILITY RPCDBG_TRANS | |
58 | #endif | |
59 | ||
60 | MODULE_LICENSE("Dual BSD/GPL"); | |
61 | ||
62 | MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS"); | |
63 | MODULE_AUTHOR("Network Appliance, Inc."); | |
64 | ||
65 | /* | |
66 | * tunables | |
67 | */ | |
68 | ||
69 | static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE; | |
70 | static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE; | |
71 | static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE; | |
72 | static unsigned int xprt_rdma_inline_write_padding; | |
3197d309 | 73 | static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR; |
9191ca3b | 74 | int xprt_rdma_pad_optimize = 0; |
f58851e6 TT |
75 | |
76 | #ifdef RPC_DEBUG | |
77 | ||
78 | static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE; | |
79 | static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE; | |
80 | static unsigned int zero; | |
81 | static unsigned int max_padding = PAGE_SIZE; | |
82 | static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS; | |
83 | static unsigned int max_memreg = RPCRDMA_LAST - 1; | |
84 | ||
85 | static struct ctl_table_header *sunrpc_table_header; | |
86 | ||
87 | static ctl_table xr_tunables_table[] = { | |
88 | { | |
f58851e6 TT |
89 | .procname = "rdma_slot_table_entries", |
90 | .data = &xprt_rdma_slot_table_entries, | |
91 | .maxlen = sizeof(unsigned int), | |
92 | .mode = 0644, | |
6d456111 | 93 | .proc_handler = proc_dointvec_minmax, |
f58851e6 TT |
94 | .extra1 = &min_slot_table_size, |
95 | .extra2 = &max_slot_table_size | |
96 | }, | |
97 | { | |
f58851e6 TT |
98 | .procname = "rdma_max_inline_read", |
99 | .data = &xprt_rdma_max_inline_read, | |
100 | .maxlen = sizeof(unsigned int), | |
101 | .mode = 0644, | |
6d456111 | 102 | .proc_handler = proc_dointvec, |
f58851e6 TT |
103 | }, |
104 | { | |
f58851e6 TT |
105 | .procname = "rdma_max_inline_write", |
106 | .data = &xprt_rdma_max_inline_write, | |
107 | .maxlen = sizeof(unsigned int), | |
108 | .mode = 0644, | |
6d456111 | 109 | .proc_handler = proc_dointvec, |
f58851e6 TT |
110 | }, |
111 | { | |
f58851e6 TT |
112 | .procname = "rdma_inline_write_padding", |
113 | .data = &xprt_rdma_inline_write_padding, | |
114 | .maxlen = sizeof(unsigned int), | |
115 | .mode = 0644, | |
6d456111 | 116 | .proc_handler = proc_dointvec_minmax, |
f58851e6 TT |
117 | .extra1 = &zero, |
118 | .extra2 = &max_padding, | |
119 | }, | |
120 | { | |
f58851e6 TT |
121 | .procname = "rdma_memreg_strategy", |
122 | .data = &xprt_rdma_memreg_strategy, | |
123 | .maxlen = sizeof(unsigned int), | |
124 | .mode = 0644, | |
6d456111 | 125 | .proc_handler = proc_dointvec_minmax, |
f58851e6 TT |
126 | .extra1 = &min_memreg, |
127 | .extra2 = &max_memreg, | |
128 | }, | |
9191ca3b | 129 | { |
9191ca3b TT |
130 | .procname = "rdma_pad_optimize", |
131 | .data = &xprt_rdma_pad_optimize, | |
132 | .maxlen = sizeof(unsigned int), | |
133 | .mode = 0644, | |
6d456111 | 134 | .proc_handler = proc_dointvec, |
9191ca3b | 135 | }, |
f8572d8f | 136 | { }, |
f58851e6 TT |
137 | }; |
138 | ||
139 | static ctl_table sunrpc_table[] = { | |
140 | { | |
f58851e6 TT |
141 | .procname = "sunrpc", |
142 | .mode = 0555, | |
143 | .child = xr_tunables_table | |
144 | }, | |
f8572d8f | 145 | { }, |
f58851e6 TT |
146 | }; |
147 | ||
148 | #endif | |
149 | ||
150 | static struct rpc_xprt_ops xprt_rdma_procs; /* forward reference */ | |
151 | ||
152 | static void | |
153 | xprt_rdma_format_addresses(struct rpc_xprt *xprt) | |
154 | { | |
c877b849 | 155 | struct sockaddr *sap = (struct sockaddr *) |
f58851e6 | 156 | &rpcx_to_rdmad(xprt).addr; |
c877b849 CL |
157 | struct sockaddr_in *sin = (struct sockaddr_in *)sap; |
158 | char buf[64]; | |
f58851e6 | 159 | |
c877b849 CL |
160 | (void)rpc_ntop(sap, buf, sizeof(buf)); |
161 | xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL); | |
f58851e6 | 162 | |
c877b849 CL |
163 | (void)snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap)); |
164 | xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL); | |
f58851e6 TT |
165 | |
166 | xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma"; | |
167 | ||
fc0b5791 | 168 | (void)snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr)); |
c877b849 CL |
169 | xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL); |
170 | ||
171 | (void)snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap)); | |
172 | xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL); | |
f58851e6 | 173 | |
f58851e6 TT |
174 | /* netid */ |
175 | xprt->address_strings[RPC_DISPLAY_NETID] = "rdma"; | |
176 | } | |
177 | ||
178 | static void | |
179 | xprt_rdma_free_addresses(struct rpc_xprt *xprt) | |
180 | { | |
33e01dc7 CL |
181 | unsigned int i; |
182 | ||
183 | for (i = 0; i < RPC_DISPLAY_MAX; i++) | |
184 | switch (i) { | |
185 | case RPC_DISPLAY_PROTO: | |
186 | case RPC_DISPLAY_NETID: | |
187 | continue; | |
188 | default: | |
189 | kfree(xprt->address_strings[i]); | |
190 | } | |
f58851e6 TT |
191 | } |
192 | ||
193 | static void | |
194 | xprt_rdma_connect_worker(struct work_struct *work) | |
195 | { | |
196 | struct rpcrdma_xprt *r_xprt = | |
197 | container_of(work, struct rpcrdma_xprt, rdma_connect.work); | |
198 | struct rpc_xprt *xprt = &r_xprt->xprt; | |
199 | int rc = 0; | |
200 | ||
201 | if (!xprt->shutdown) { | |
202 | xprt_clear_connected(xprt); | |
203 | ||
204 | dprintk("RPC: %s: %sconnect\n", __func__, | |
205 | r_xprt->rx_ep.rep_connected != 0 ? "re" : ""); | |
206 | rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia); | |
207 | if (rc) | |
208 | goto out; | |
209 | } | |
210 | goto out_clear; | |
211 | ||
212 | out: | |
213 | xprt_wake_pending_tasks(xprt, rc); | |
214 | ||
215 | out_clear: | |
216 | dprintk("RPC: %s: exit\n", __func__); | |
217 | xprt_clear_connecting(xprt); | |
218 | } | |
219 | ||
220 | /* | |
221 | * xprt_rdma_destroy | |
222 | * | |
223 | * Destroy the xprt. | |
224 | * Free all memory associated with the object, including its own. | |
225 | * NOTE: none of the *destroy methods free memory for their top-level | |
226 | * objects, even though they may have allocated it (they do free | |
227 | * private memory). It's up to the caller to handle it. In this | |
228 | * case (RDMA transport), all structure memory is inlined with the | |
229 | * struct rpcrdma_xprt. | |
230 | */ | |
231 | static void | |
232 | xprt_rdma_destroy(struct rpc_xprt *xprt) | |
233 | { | |
234 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); | |
235 | int rc; | |
236 | ||
237 | dprintk("RPC: %s: called\n", __func__); | |
238 | ||
239 | cancel_delayed_work(&r_xprt->rdma_connect); | |
240 | flush_scheduled_work(); | |
241 | ||
242 | xprt_clear_connected(xprt); | |
243 | ||
244 | rpcrdma_buffer_destroy(&r_xprt->rx_buf); | |
245 | rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia); | |
246 | if (rc) | |
247 | dprintk("RPC: %s: rpcrdma_ep_destroy returned %i\n", | |
248 | __func__, rc); | |
249 | rpcrdma_ia_close(&r_xprt->rx_ia); | |
250 | ||
251 | xprt_rdma_free_addresses(xprt); | |
252 | ||
253 | kfree(xprt->slot); | |
254 | xprt->slot = NULL; | |
255 | kfree(xprt); | |
256 | ||
257 | dprintk("RPC: %s: returning\n", __func__); | |
258 | ||
259 | module_put(THIS_MODULE); | |
260 | } | |
261 | ||
2881ae74 TM |
262 | static const struct rpc_timeout xprt_rdma_default_timeout = { |
263 | .to_initval = 60 * HZ, | |
264 | .to_maxval = 60 * HZ, | |
265 | }; | |
266 | ||
f58851e6 TT |
267 | /** |
268 | * xprt_setup_rdma - Set up transport to use RDMA | |
269 | * | |
270 | * @args: rpc transport arguments | |
271 | */ | |
272 | static struct rpc_xprt * | |
273 | xprt_setup_rdma(struct xprt_create *args) | |
274 | { | |
275 | struct rpcrdma_create_data_internal cdata; | |
276 | struct rpc_xprt *xprt; | |
277 | struct rpcrdma_xprt *new_xprt; | |
278 | struct rpcrdma_ep *new_ep; | |
279 | struct sockaddr_in *sin; | |
280 | int rc; | |
281 | ||
282 | if (args->addrlen > sizeof(xprt->addr)) { | |
283 | dprintk("RPC: %s: address too large\n", __func__); | |
284 | return ERR_PTR(-EBADF); | |
285 | } | |
286 | ||
287 | xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL); | |
288 | if (xprt == NULL) { | |
289 | dprintk("RPC: %s: couldn't allocate rpcrdma_xprt\n", | |
290 | __func__); | |
291 | return ERR_PTR(-ENOMEM); | |
292 | } | |
293 | ||
294 | xprt->max_reqs = xprt_rdma_slot_table_entries; | |
295 | xprt->slot = kcalloc(xprt->max_reqs, | |
296 | sizeof(struct rpc_rqst), GFP_KERNEL); | |
297 | if (xprt->slot == NULL) { | |
f58851e6 TT |
298 | dprintk("RPC: %s: couldn't allocate %d slots\n", |
299 | __func__, xprt->max_reqs); | |
d5cd9787 | 300 | kfree(xprt); |
f58851e6 TT |
301 | return ERR_PTR(-ENOMEM); |
302 | } | |
303 | ||
304 | /* 60 second timeout, no retries */ | |
ba7392bb | 305 | xprt->timeout = &xprt_rdma_default_timeout; |
f58851e6 TT |
306 | xprt->bind_timeout = (60U * HZ); |
307 | xprt->connect_timeout = (60U * HZ); | |
308 | xprt->reestablish_timeout = (5U * HZ); | |
309 | xprt->idle_timeout = (5U * 60 * HZ); | |
310 | ||
311 | xprt->resvport = 0; /* privileged port not needed */ | |
312 | xprt->tsh_size = 0; /* RPC-RDMA handles framing */ | |
313 | xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE; | |
314 | xprt->ops = &xprt_rdma_procs; | |
315 | ||
316 | /* | |
317 | * Set up RDMA-specific connect data. | |
318 | */ | |
319 | ||
320 | /* Put server RDMA address in local cdata */ | |
321 | memcpy(&cdata.addr, args->dstaddr, args->addrlen); | |
322 | ||
323 | /* Ensure xprt->addr holds valid server TCP (not RDMA) | |
324 | * address, for any side protocols which peek at it */ | |
325 | xprt->prot = IPPROTO_TCP; | |
326 | xprt->addrlen = args->addrlen; | |
327 | memcpy(&xprt->addr, &cdata.addr, xprt->addrlen); | |
328 | ||
329 | sin = (struct sockaddr_in *)&cdata.addr; | |
330 | if (ntohs(sin->sin_port) != 0) | |
331 | xprt_set_bound(xprt); | |
332 | ||
21454aaa HH |
333 | dprintk("RPC: %s: %pI4:%u\n", |
334 | __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port)); | |
f58851e6 TT |
335 | |
336 | /* Set max requests */ | |
337 | cdata.max_requests = xprt->max_reqs; | |
338 | ||
339 | /* Set some length limits */ | |
340 | cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */ | |
341 | cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */ | |
342 | ||
343 | cdata.inline_wsize = xprt_rdma_max_inline_write; | |
344 | if (cdata.inline_wsize > cdata.wsize) | |
345 | cdata.inline_wsize = cdata.wsize; | |
346 | ||
347 | cdata.inline_rsize = xprt_rdma_max_inline_read; | |
348 | if (cdata.inline_rsize > cdata.rsize) | |
349 | cdata.inline_rsize = cdata.rsize; | |
350 | ||
351 | cdata.padding = xprt_rdma_inline_write_padding; | |
352 | ||
353 | /* | |
354 | * Create new transport instance, which includes initialized | |
355 | * o ia | |
356 | * o endpoint | |
357 | * o buffers | |
358 | */ | |
359 | ||
360 | new_xprt = rpcx_to_rdmax(xprt); | |
361 | ||
362 | rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr, | |
363 | xprt_rdma_memreg_strategy); | |
364 | if (rc) | |
365 | goto out1; | |
366 | ||
367 | /* | |
368 | * initialize and create ep | |
369 | */ | |
370 | new_xprt->rx_data = cdata; | |
371 | new_ep = &new_xprt->rx_ep; | |
372 | new_ep->rep_remote_addr = cdata.addr; | |
373 | ||
374 | rc = rpcrdma_ep_create(&new_xprt->rx_ep, | |
375 | &new_xprt->rx_ia, &new_xprt->rx_data); | |
376 | if (rc) | |
377 | goto out2; | |
378 | ||
379 | /* | |
380 | * Allocate pre-registered send and receive buffers for headers and | |
381 | * any inline data. Also specify any padding which will be provided | |
382 | * from a preregistered zero buffer. | |
383 | */ | |
384 | rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia, | |
385 | &new_xprt->rx_data); | |
386 | if (rc) | |
387 | goto out3; | |
388 | ||
389 | /* | |
390 | * Register a callback for connection events. This is necessary because | |
391 | * connection loss notification is async. We also catch connection loss | |
392 | * when reaping receives. | |
393 | */ | |
394 | INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker); | |
395 | new_ep->rep_func = rpcrdma_conn_func; | |
396 | new_ep->rep_xprt = xprt; | |
397 | ||
398 | xprt_rdma_format_addresses(xprt); | |
399 | ||
400 | if (!try_module_get(THIS_MODULE)) | |
401 | goto out4; | |
402 | ||
403 | return xprt; | |
404 | ||
405 | out4: | |
406 | xprt_rdma_free_addresses(xprt); | |
407 | rc = -EINVAL; | |
408 | out3: | |
409 | (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia); | |
410 | out2: | |
411 | rpcrdma_ia_close(&new_xprt->rx_ia); | |
412 | out1: | |
413 | kfree(xprt->slot); | |
414 | kfree(xprt); | |
415 | return ERR_PTR(rc); | |
416 | } | |
417 | ||
418 | /* | |
419 | * Close a connection, during shutdown or timeout/reconnect | |
420 | */ | |
421 | static void | |
422 | xprt_rdma_close(struct rpc_xprt *xprt) | |
423 | { | |
424 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); | |
425 | ||
426 | dprintk("RPC: %s: closing\n", __func__); | |
08ca0dce TT |
427 | if (r_xprt->rx_ep.rep_connected > 0) |
428 | xprt->reestablish_timeout = 0; | |
62da3b24 | 429 | xprt_disconnect_done(xprt); |
f58851e6 TT |
430 | (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia); |
431 | } | |
432 | ||
433 | static void | |
434 | xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port) | |
435 | { | |
436 | struct sockaddr_in *sap; | |
437 | ||
438 | sap = (struct sockaddr_in *)&xprt->addr; | |
439 | sap->sin_port = htons(port); | |
440 | sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr; | |
441 | sap->sin_port = htons(port); | |
442 | dprintk("RPC: %s: %u\n", __func__, port); | |
443 | } | |
444 | ||
445 | static void | |
446 | xprt_rdma_connect(struct rpc_task *task) | |
447 | { | |
448 | struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt; | |
449 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); | |
450 | ||
451 | if (!xprt_test_and_set_connecting(xprt)) { | |
452 | if (r_xprt->rx_ep.rep_connected != 0) { | |
453 | /* Reconnect */ | |
454 | schedule_delayed_work(&r_xprt->rdma_connect, | |
455 | xprt->reestablish_timeout); | |
08ca0dce TT |
456 | xprt->reestablish_timeout <<= 1; |
457 | if (xprt->reestablish_timeout > (30 * HZ)) | |
458 | xprt->reestablish_timeout = (30 * HZ); | |
459 | else if (xprt->reestablish_timeout < (5 * HZ)) | |
460 | xprt->reestablish_timeout = (5 * HZ); | |
f58851e6 TT |
461 | } else { |
462 | schedule_delayed_work(&r_xprt->rdma_connect, 0); | |
463 | if (!RPC_IS_ASYNC(task)) | |
464 | flush_scheduled_work(); | |
465 | } | |
466 | } | |
467 | } | |
468 | ||
469 | static int | |
470 | xprt_rdma_reserve_xprt(struct rpc_task *task) | |
471 | { | |
472 | struct rpc_xprt *xprt = task->tk_xprt; | |
473 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); | |
474 | int credits = atomic_read(&r_xprt->rx_buf.rb_credits); | |
475 | ||
476 | /* == RPC_CWNDSCALE @ init, but *after* setup */ | |
477 | if (r_xprt->rx_buf.rb_cwndscale == 0UL) { | |
478 | r_xprt->rx_buf.rb_cwndscale = xprt->cwnd; | |
479 | dprintk("RPC: %s: cwndscale %lu\n", __func__, | |
480 | r_xprt->rx_buf.rb_cwndscale); | |
481 | BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0); | |
482 | } | |
483 | xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale; | |
484 | return xprt_reserve_xprt_cong(task); | |
485 | } | |
486 | ||
487 | /* | |
488 | * The RDMA allocate/free functions need the task structure as a place | |
489 | * to hide the struct rpcrdma_req, which is necessary for the actual send/recv | |
490 | * sequence. For this reason, the recv buffers are attached to send | |
491 | * buffers for portions of the RPC. Note that the RPC layer allocates | |
492 | * both send and receive buffers in the same call. We may register | |
493 | * the receive buffer portion when using reply chunks. | |
494 | */ | |
495 | static void * | |
496 | xprt_rdma_allocate(struct rpc_task *task, size_t size) | |
497 | { | |
498 | struct rpc_xprt *xprt = task->tk_xprt; | |
499 | struct rpcrdma_req *req, *nreq; | |
500 | ||
501 | req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf); | |
502 | BUG_ON(NULL == req); | |
503 | ||
504 | if (size > req->rl_size) { | |
505 | dprintk("RPC: %s: size %zd too large for buffer[%zd]: " | |
506 | "prog %d vers %d proc %d\n", | |
507 | __func__, size, req->rl_size, | |
508 | task->tk_client->cl_prog, task->tk_client->cl_vers, | |
509 | task->tk_msg.rpc_proc->p_proc); | |
510 | /* | |
511 | * Outgoing length shortage. Our inline write max must have | |
512 | * been configured to perform direct i/o. | |
513 | * | |
514 | * This is therefore a large metadata operation, and the | |
515 | * allocate call was made on the maximum possible message, | |
516 | * e.g. containing long filename(s) or symlink data. In | |
517 | * fact, while these metadata operations *might* carry | |
518 | * large outgoing payloads, they rarely *do*. However, we | |
519 | * have to commit to the request here, so reallocate and | |
520 | * register it now. The data path will never require this | |
521 | * reallocation. | |
522 | * | |
523 | * If the allocation or registration fails, the RPC framework | |
524 | * will (doggedly) retry. | |
525 | */ | |
526 | if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy == | |
527 | RPCRDMA_BOUNCEBUFFERS) { | |
528 | /* forced to "pure inline" */ | |
529 | dprintk("RPC: %s: too much data (%zd) for inline " | |
530 | "(r/w max %d/%d)\n", __func__, size, | |
531 | rpcx_to_rdmad(xprt).inline_rsize, | |
532 | rpcx_to_rdmad(xprt).inline_wsize); | |
533 | size = req->rl_size; | |
534 | rpc_exit(task, -EIO); /* fail the operation */ | |
535 | rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++; | |
536 | goto out; | |
537 | } | |
538 | if (task->tk_flags & RPC_TASK_SWAPPER) | |
539 | nreq = kmalloc(sizeof *req + size, GFP_ATOMIC); | |
540 | else | |
541 | nreq = kmalloc(sizeof *req + size, GFP_NOFS); | |
542 | if (nreq == NULL) | |
543 | goto outfail; | |
544 | ||
545 | if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia, | |
546 | nreq->rl_base, size + sizeof(struct rpcrdma_req) | |
547 | - offsetof(struct rpcrdma_req, rl_base), | |
548 | &nreq->rl_handle, &nreq->rl_iov)) { | |
549 | kfree(nreq); | |
550 | goto outfail; | |
551 | } | |
552 | rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size; | |
553 | nreq->rl_size = size; | |
554 | nreq->rl_niovs = 0; | |
555 | nreq->rl_nchunks = 0; | |
556 | nreq->rl_buffer = (struct rpcrdma_buffer *)req; | |
557 | nreq->rl_reply = req->rl_reply; | |
558 | memcpy(nreq->rl_segments, | |
559 | req->rl_segments, sizeof nreq->rl_segments); | |
560 | /* flag the swap with an unused field */ | |
561 | nreq->rl_iov.length = 0; | |
562 | req->rl_reply = NULL; | |
563 | req = nreq; | |
564 | } | |
565 | dprintk("RPC: %s: size %zd, request 0x%p\n", __func__, size, req); | |
566 | out: | |
575448bd | 567 | req->rl_connect_cookie = 0; /* our reserved value */ |
f58851e6 TT |
568 | return req->rl_xdr_buf; |
569 | ||
570 | outfail: | |
571 | rpcrdma_buffer_put(req); | |
572 | rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++; | |
573 | return NULL; | |
574 | } | |
575 | ||
576 | /* | |
577 | * This function returns all RDMA resources to the pool. | |
578 | */ | |
579 | static void | |
580 | xprt_rdma_free(void *buffer) | |
581 | { | |
582 | struct rpcrdma_req *req; | |
583 | struct rpcrdma_xprt *r_xprt; | |
584 | struct rpcrdma_rep *rep; | |
585 | int i; | |
586 | ||
587 | if (buffer == NULL) | |
588 | return; | |
589 | ||
590 | req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]); | |
ee1a2c56 TT |
591 | if (req->rl_iov.length == 0) { /* see allocate above */ |
592 | r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer, | |
593 | struct rpcrdma_xprt, rx_buf); | |
594 | } else | |
595 | r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf); | |
f58851e6 TT |
596 | rep = req->rl_reply; |
597 | ||
598 | dprintk("RPC: %s: called on 0x%p%s\n", | |
599 | __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : ""); | |
600 | ||
601 | /* | |
602 | * Finish the deregistration. When using mw bind, this was | |
603 | * begun in rpcrdma_reply_handler(). In all other modes, we | |
604 | * do it here, in thread context. The process is considered | |
605 | * complete when the rr_func vector becomes NULL - this | |
606 | * was put in place during rpcrdma_reply_handler() - the wait | |
607 | * call below will not block if the dereg is "done". If | |
608 | * interrupted, our framework will clean up. | |
609 | */ | |
610 | for (i = 0; req->rl_nchunks;) { | |
611 | --req->rl_nchunks; | |
612 | i += rpcrdma_deregister_external( | |
613 | &req->rl_segments[i], r_xprt, NULL); | |
614 | } | |
615 | ||
616 | if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) { | |
617 | rep->rr_func = NULL; /* abandon the callback */ | |
618 | req->rl_reply = NULL; | |
619 | } | |
620 | ||
621 | if (req->rl_iov.length == 0) { /* see allocate above */ | |
622 | struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer; | |
623 | oreq->rl_reply = req->rl_reply; | |
624 | (void) rpcrdma_deregister_internal(&r_xprt->rx_ia, | |
625 | req->rl_handle, | |
626 | &req->rl_iov); | |
627 | kfree(req); | |
628 | req = oreq; | |
629 | } | |
630 | ||
631 | /* Put back request+reply buffers */ | |
632 | rpcrdma_buffer_put(req); | |
633 | } | |
634 | ||
635 | /* | |
636 | * send_request invokes the meat of RPC RDMA. It must do the following: | |
637 | * 1. Marshal the RPC request into an RPC RDMA request, which means | |
638 | * putting a header in front of data, and creating IOVs for RDMA | |
639 | * from those in the request. | |
640 | * 2. In marshaling, detect opportunities for RDMA, and use them. | |
641 | * 3. Post a recv message to set up asynch completion, then send | |
642 | * the request (rpcrdma_ep_post). | |
643 | * 4. No partial sends are possible in the RPC-RDMA protocol (as in UDP). | |
644 | */ | |
645 | ||
646 | static int | |
647 | xprt_rdma_send_request(struct rpc_task *task) | |
648 | { | |
649 | struct rpc_rqst *rqst = task->tk_rqstp; | |
650 | struct rpc_xprt *xprt = task->tk_xprt; | |
651 | struct rpcrdma_req *req = rpcr_to_rdmar(rqst); | |
652 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); | |
653 | ||
654 | /* marshal the send itself */ | |
655 | if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) { | |
656 | r_xprt->rx_stats.failed_marshal_count++; | |
657 | dprintk("RPC: %s: rpcrdma_marshal_req failed\n", | |
658 | __func__); | |
659 | return -EIO; | |
660 | } | |
661 | ||
662 | if (req->rl_reply == NULL) /* e.g. reconnection */ | |
663 | rpcrdma_recv_buffer_get(req); | |
664 | ||
665 | if (req->rl_reply) { | |
666 | req->rl_reply->rr_func = rpcrdma_reply_handler; | |
667 | /* this need only be done once, but... */ | |
668 | req->rl_reply->rr_xprt = xprt; | |
669 | } | |
670 | ||
575448bd TT |
671 | /* Must suppress retransmit to maintain credits */ |
672 | if (req->rl_connect_cookie == xprt->connect_cookie) | |
673 | goto drop_connection; | |
674 | req->rl_connect_cookie = xprt->connect_cookie; | |
675 | ||
676 | if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req)) | |
677 | goto drop_connection; | |
f58851e6 | 678 | |
ad0e9e01 | 679 | task->tk_bytes_sent += rqst->rq_snd_buf.len; |
f58851e6 TT |
680 | rqst->rq_bytes_sent = 0; |
681 | return 0; | |
575448bd TT |
682 | |
683 | drop_connection: | |
684 | xprt_disconnect_done(xprt); | |
685 | return -ENOTCONN; /* implies disconnect */ | |
f58851e6 TT |
686 | } |
687 | ||
688 | static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq) | |
689 | { | |
690 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); | |
691 | long idle_time = 0; | |
692 | ||
693 | if (xprt_connected(xprt)) | |
694 | idle_time = (long)(jiffies - xprt->last_used) / HZ; | |
695 | ||
696 | seq_printf(seq, | |
697 | "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu " | |
698 | "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n", | |
699 | ||
700 | 0, /* need a local port? */ | |
701 | xprt->stat.bind_count, | |
702 | xprt->stat.connect_count, | |
703 | xprt->stat.connect_time, | |
704 | idle_time, | |
705 | xprt->stat.sends, | |
706 | xprt->stat.recvs, | |
707 | xprt->stat.bad_xids, | |
708 | xprt->stat.req_u, | |
709 | xprt->stat.bklog_u, | |
710 | ||
711 | r_xprt->rx_stats.read_chunk_count, | |
712 | r_xprt->rx_stats.write_chunk_count, | |
713 | r_xprt->rx_stats.reply_chunk_count, | |
714 | r_xprt->rx_stats.total_rdma_request, | |
715 | r_xprt->rx_stats.total_rdma_reply, | |
716 | r_xprt->rx_stats.pullup_copy_count, | |
717 | r_xprt->rx_stats.fixup_copy_count, | |
718 | r_xprt->rx_stats.hardway_register_count, | |
719 | r_xprt->rx_stats.failed_marshal_count, | |
720 | r_xprt->rx_stats.bad_reply_count); | |
721 | } | |
722 | ||
723 | /* | |
724 | * Plumbing for rpc transport switch and kernel module | |
725 | */ | |
726 | ||
727 | static struct rpc_xprt_ops xprt_rdma_procs = { | |
728 | .reserve_xprt = xprt_rdma_reserve_xprt, | |
729 | .release_xprt = xprt_release_xprt_cong, /* sunrpc/xprt.c */ | |
730 | .release_request = xprt_release_rqst_cong, /* ditto */ | |
731 | .set_retrans_timeout = xprt_set_retrans_timeout_def, /* ditto */ | |
732 | .rpcbind = rpcb_getport_async, /* sunrpc/rpcb_clnt.c */ | |
733 | .set_port = xprt_rdma_set_port, | |
734 | .connect = xprt_rdma_connect, | |
735 | .buf_alloc = xprt_rdma_allocate, | |
736 | .buf_free = xprt_rdma_free, | |
737 | .send_request = xprt_rdma_send_request, | |
738 | .close = xprt_rdma_close, | |
739 | .destroy = xprt_rdma_destroy, | |
740 | .print_stats = xprt_rdma_print_stats | |
741 | }; | |
742 | ||
743 | static struct xprt_class xprt_rdma = { | |
744 | .list = LIST_HEAD_INIT(xprt_rdma.list), | |
745 | .name = "rdma", | |
746 | .owner = THIS_MODULE, | |
747 | .ident = XPRT_TRANSPORT_RDMA, | |
748 | .setup = xprt_setup_rdma, | |
749 | }; | |
750 | ||
751 | static void __exit xprt_rdma_cleanup(void) | |
752 | { | |
753 | int rc; | |
754 | ||
b3cd8d45 | 755 | dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n"); |
f58851e6 TT |
756 | #ifdef RPC_DEBUG |
757 | if (sunrpc_table_header) { | |
758 | unregister_sysctl_table(sunrpc_table_header); | |
759 | sunrpc_table_header = NULL; | |
760 | } | |
761 | #endif | |
762 | rc = xprt_unregister_transport(&xprt_rdma); | |
763 | if (rc) | |
764 | dprintk("RPC: %s: xprt_unregister returned %i\n", | |
765 | __func__, rc); | |
766 | } | |
767 | ||
768 | static int __init xprt_rdma_init(void) | |
769 | { | |
770 | int rc; | |
771 | ||
772 | rc = xprt_register_transport(&xprt_rdma); | |
773 | ||
774 | if (rc) | |
775 | return rc; | |
776 | ||
777 | dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n"); | |
778 | ||
779 | dprintk(KERN_INFO "Defaults:\n"); | |
780 | dprintk(KERN_INFO "\tSlots %d\n" | |
781 | "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n", | |
782 | xprt_rdma_slot_table_entries, | |
783 | xprt_rdma_max_inline_read, xprt_rdma_max_inline_write); | |
784 | dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n", | |
785 | xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy); | |
786 | ||
787 | #ifdef RPC_DEBUG | |
788 | if (!sunrpc_table_header) | |
789 | sunrpc_table_header = register_sysctl_table(sunrpc_table); | |
790 | #endif | |
791 | return 0; | |
792 | } | |
793 | ||
794 | module_init(xprt_rdma_init); | |
795 | module_exit(xprt_rdma_cleanup); |