]>
Commit | Line | Data |
---|---|---|
7ab672ce DH |
1 | /* |
2 | * Module for the pnfs nfs4 file layout driver. | |
3 | * Defines all I/O and Policy interface operations, plus code | |
4 | * to register itself with the pNFS client. | |
5 | * | |
6 | * Copyright (c) 2002 | |
7 | * The Regents of the University of Michigan | |
8 | * All Rights Reserved | |
9 | * | |
10 | * Dean Hildebrand <dhildebz@umich.edu> | |
11 | * | |
12 | * Permission is granted to use, copy, create derivative works, and | |
13 | * redistribute this software and such derivative works for any purpose, | |
14 | * so long as the name of the University of Michigan is not used in | |
15 | * any advertising or publicity pertaining to the use or distribution | |
16 | * of this software without specific, written prior authorization. If | |
17 | * the above copyright notice or any other identification of the | |
18 | * University of Michigan is included in any copy of any portion of | |
19 | * this software, then the disclaimer below must also be included. | |
20 | * | |
21 | * This software is provided as is, without representation or warranty | |
22 | * of any kind either express or implied, including without limitation | |
23 | * the implied warranties of merchantability, fitness for a particular | |
24 | * purpose, or noninfringement. The Regents of the University of | |
25 | * Michigan shall not be liable for any damages, including special, | |
26 | * indirect, incidental, or consequential damages, with respect to any | |
27 | * claim arising out of or in connection with the use of the software, | |
28 | * even if it has been or is hereafter advised of the possibility of | |
29 | * such damages. | |
30 | */ | |
31 | ||
32 | #include <linux/nfs_fs.h> | |
19345cb2 | 33 | #include <linux/nfs_page.h> |
143cb494 | 34 | #include <linux/module.h> |
16b374ca | 35 | |
0a702195 WAA |
36 | #include <linux/sunrpc/metrics.h> |
37 | ||
16b374ca | 38 | #include "internal.h" |
9cb81968 | 39 | #include "delegation.h" |
16b374ca | 40 | #include "nfs4filelayout.h" |
7ab672ce DH |
41 | |
42 | #define NFSDBG_FACILITY NFSDBG_PNFS_LD | |
43 | ||
44 | MODULE_LICENSE("GPL"); | |
45 | MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>"); | |
46 | MODULE_DESCRIPTION("The NFSv4 file layout driver"); | |
47 | ||
cbdabc7f AA |
48 | #define FILELAYOUT_POLL_RETRY_MAX (15*HZ) |
49 | ||
cfe7f412 FI |
50 | static loff_t |
51 | filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg, | |
52 | loff_t offset) | |
53 | { | |
54 | u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count; | |
3476f114 CM |
55 | u64 stripe_no; |
56 | u32 rem; | |
cfe7f412 FI |
57 | |
58 | offset -= flseg->pattern_offset; | |
3476f114 CM |
59 | stripe_no = div_u64(offset, stripe_width); |
60 | div_u64_rem(offset, flseg->stripe_unit, &rem); | |
cfe7f412 | 61 | |
3476f114 | 62 | return stripe_no * flseg->stripe_unit + rem; |
cfe7f412 FI |
63 | } |
64 | ||
65 | /* This function is used by the layout driver to calculate the | |
66 | * offset of the file on the dserver based on whether the | |
67 | * layout type is STRIPE_DENSE or STRIPE_SPARSE | |
68 | */ | |
69 | static loff_t | |
70 | filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset) | |
71 | { | |
72 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
73 | ||
74 | switch (flseg->stripe_type) { | |
75 | case STRIPE_SPARSE: | |
76 | return offset; | |
77 | ||
78 | case STRIPE_DENSE: | |
79 | return filelayout_get_dense_offset(flseg, offset); | |
80 | } | |
81 | ||
82 | BUG(); | |
83 | } | |
84 | ||
e7dd79af AA |
85 | static void filelayout_reset_write(struct nfs_write_data *data) |
86 | { | |
87 | struct nfs_pgio_header *hdr = data->header; | |
88 | struct inode *inode = hdr->inode; | |
89 | struct rpc_task *task = &data->task; | |
90 | ||
91 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { | |
92 | dprintk("%s Reset task %5u for i/o through MDS " | |
93 | "(req %s/%lld, %u bytes @ offset %llu)\n", __func__, | |
94 | data->task.tk_pid, | |
95 | inode->i_sb->s_id, | |
96 | (long long)NFS_FILEID(inode), | |
97 | data->args.count, | |
98 | (unsigned long long)data->args.offset); | |
99 | ||
100 | task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode, | |
101 | &hdr->pages, | |
102 | hdr->completion_ops); | |
103 | } | |
104 | } | |
105 | ||
106 | static void filelayout_reset_read(struct nfs_read_data *data) | |
107 | { | |
108 | struct nfs_pgio_header *hdr = data->header; | |
109 | struct inode *inode = hdr->inode; | |
110 | struct rpc_task *task = &data->task; | |
111 | ||
112 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { | |
113 | dprintk("%s Reset task %5u for i/o through MDS " | |
114 | "(req %s/%lld, %u bytes @ offset %llu)\n", __func__, | |
115 | data->task.tk_pid, | |
116 | inode->i_sb->s_id, | |
117 | (long long)NFS_FILEID(inode), | |
118 | data->args.count, | |
119 | (unsigned long long)data->args.offset); | |
120 | ||
121 | task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode, | |
122 | &hdr->pages, | |
123 | hdr->completion_ops); | |
124 | } | |
125 | } | |
126 | ||
cbdabc7f AA |
127 | static int filelayout_async_handle_error(struct rpc_task *task, |
128 | struct nfs4_state *state, | |
129 | struct nfs_client *clp, | |
e7dd79af | 130 | struct pnfs_layout_segment *lseg) |
cbdabc7f | 131 | { |
e7dd79af AA |
132 | struct inode *inode = lseg->pls_layout->plh_inode; |
133 | struct nfs_server *mds_server = NFS_SERVER(inode); | |
134 | struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg); | |
9cb81968 AA |
135 | struct nfs_client *mds_client = mds_server->nfs_client; |
136 | ||
cbdabc7f AA |
137 | if (task->tk_status >= 0) |
138 | return 0; | |
cbdabc7f AA |
139 | |
140 | switch (task->tk_status) { | |
9cb81968 AA |
141 | /* MDS state errors */ |
142 | case -NFS4ERR_DELEG_REVOKED: | |
143 | case -NFS4ERR_ADMIN_REVOKED: | |
144 | case -NFS4ERR_BAD_STATEID: | |
e7dd79af AA |
145 | if (state == NULL) |
146 | break; | |
2dc31756 | 147 | nfs_remove_bad_delegation(state->inode); |
9cb81968 | 148 | case -NFS4ERR_OPENMODE: |
e7dd79af AA |
149 | if (state == NULL) |
150 | break; | |
9cb81968 AA |
151 | nfs4_schedule_stateid_recovery(mds_server, state); |
152 | goto wait_on_recovery; | |
153 | case -NFS4ERR_EXPIRED: | |
e7dd79af AA |
154 | if (state != NULL) |
155 | nfs4_schedule_stateid_recovery(mds_server, state); | |
9cb81968 AA |
156 | nfs4_schedule_lease_recovery(mds_client); |
157 | goto wait_on_recovery; | |
158 | /* DS session errors */ | |
cbdabc7f AA |
159 | case -NFS4ERR_BADSESSION: |
160 | case -NFS4ERR_BADSLOT: | |
161 | case -NFS4ERR_BAD_HIGH_SLOT: | |
162 | case -NFS4ERR_DEADSESSION: | |
163 | case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION: | |
164 | case -NFS4ERR_SEQ_FALSE_RETRY: | |
165 | case -NFS4ERR_SEQ_MISORDERED: | |
166 | dprintk("%s ERROR %d, Reset session. Exchangeid " | |
167 | "flags 0x%x\n", __func__, task->tk_status, | |
168 | clp->cl_exchange_flags); | |
169 | nfs4_schedule_session_recovery(clp->cl_session); | |
170 | break; | |
171 | case -NFS4ERR_DELAY: | |
172 | case -NFS4ERR_GRACE: | |
173 | case -EKEYEXPIRED: | |
174 | rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX); | |
175 | break; | |
a8a4ae3a AA |
176 | case -NFS4ERR_RETRY_UNCACHED_REP: |
177 | break; | |
e7dd79af AA |
178 | /* RPC connection errors */ |
179 | case -ECONNREFUSED: | |
180 | case -EHOSTDOWN: | |
181 | case -EHOSTUNREACH: | |
182 | case -ENETUNREACH: | |
183 | case -EIO: | |
184 | case -ETIMEDOUT: | |
185 | case -EPIPE: | |
186 | dprintk("%s DS connection error %d\n", __func__, | |
187 | task->tk_status); | |
188 | filelayout_mark_devid_invalid(devid); | |
189 | /* fall through */ | |
cbdabc7f | 190 | default: |
e7dd79af | 191 | dprintk("%s Retry through MDS. Error %d\n", __func__, |
cbdabc7f | 192 | task->tk_status); |
e7dd79af | 193 | return -NFS4ERR_RESET_TO_MDS; |
cbdabc7f | 194 | } |
9cb81968 | 195 | out: |
cbdabc7f AA |
196 | task->tk_status = 0; |
197 | return -EAGAIN; | |
9cb81968 AA |
198 | wait_on_recovery: |
199 | rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL); | |
200 | if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0) | |
201 | rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task); | |
202 | goto out; | |
cbdabc7f AA |
203 | } |
204 | ||
205 | /* NFS_PROTO call done callback routines */ | |
206 | ||
207 | static int filelayout_read_done_cb(struct rpc_task *task, | |
208 | struct nfs_read_data *data) | |
209 | { | |
e7dd79af AA |
210 | struct nfs_pgio_header *hdr = data->header; |
211 | int err; | |
cbdabc7f | 212 | |
e7dd79af AA |
213 | err = filelayout_async_handle_error(task, data->args.context->state, |
214 | data->ds_clp, hdr->lseg); | |
cbdabc7f | 215 | |
e7dd79af AA |
216 | switch (err) { |
217 | case -NFS4ERR_RESET_TO_MDS: | |
218 | filelayout_reset_read(data); | |
219 | return task->tk_status; | |
220 | case -EAGAIN: | |
d00c5d43 | 221 | rpc_restart_call_prepare(task); |
cbdabc7f AA |
222 | return -EAGAIN; |
223 | } | |
224 | ||
225 | return 0; | |
226 | } | |
227 | ||
863a3c6c AA |
228 | /* |
229 | * We reference the rpc_cred of the first WRITE that triggers the need for | |
230 | * a LAYOUTCOMMIT, and use it to send the layoutcommit compound. | |
231 | * rfc5661 is not clear about which credential should be used. | |
232 | */ | |
233 | static void | |
234 | filelayout_set_layoutcommit(struct nfs_write_data *wdata) | |
235 | { | |
cd841605 FI |
236 | struct nfs_pgio_header *hdr = wdata->header; |
237 | ||
238 | if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds || | |
863a3c6c AA |
239 | wdata->res.verf->committed == NFS_FILE_SYNC) |
240 | return; | |
241 | ||
242 | pnfs_set_layoutcommit(wdata); | |
cd841605 FI |
243 | dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino, |
244 | (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb); | |
863a3c6c AA |
245 | } |
246 | ||
dc70d7b3 AA |
247 | /* |
248 | * Call ops for the async read/write cases | |
249 | * In the case of dense layouts, the offset needs to be reset to its | |
250 | * original value. | |
251 | */ | |
252 | static void filelayout_read_prepare(struct rpc_task *task, void *data) | |
253 | { | |
cd12ae32 | 254 | struct nfs_read_data *rdata = data; |
dc70d7b3 | 255 | |
cbdabc7f AA |
256 | rdata->read_done_cb = filelayout_read_done_cb; |
257 | ||
dc70d7b3 AA |
258 | if (nfs41_setup_sequence(rdata->ds_clp->cl_session, |
259 | &rdata->args.seq_args, &rdata->res.seq_res, | |
9d12b216 | 260 | task)) |
dc70d7b3 AA |
261 | return; |
262 | ||
263 | rpc_call_start(task); | |
264 | } | |
265 | ||
266 | static void filelayout_read_call_done(struct rpc_task *task, void *data) | |
267 | { | |
cd12ae32 | 268 | struct nfs_read_data *rdata = data; |
dc70d7b3 AA |
269 | |
270 | dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status); | |
271 | ||
272 | /* Note this may cause RPC to be resent */ | |
cd841605 | 273 | rdata->header->mds_ops->rpc_call_done(task, data); |
dc70d7b3 AA |
274 | } |
275 | ||
0a702195 WAA |
276 | static void filelayout_read_count_stats(struct rpc_task *task, void *data) |
277 | { | |
cd12ae32 | 278 | struct nfs_read_data *rdata = data; |
0a702195 | 279 | |
cd841605 | 280 | rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics); |
0a702195 WAA |
281 | } |
282 | ||
dc70d7b3 AA |
283 | static void filelayout_read_release(void *data) |
284 | { | |
cd12ae32 | 285 | struct nfs_read_data *rdata = data; |
dc70d7b3 | 286 | |
cd841605 | 287 | rdata->header->mds_ops->rpc_release(data); |
dc70d7b3 AA |
288 | } |
289 | ||
a69aef14 FI |
290 | static int filelayout_write_done_cb(struct rpc_task *task, |
291 | struct nfs_write_data *data) | |
292 | { | |
e7dd79af AA |
293 | struct nfs_pgio_header *hdr = data->header; |
294 | int err; | |
295 | ||
296 | err = filelayout_async_handle_error(task, data->args.context->state, | |
297 | data->ds_clp, hdr->lseg); | |
298 | ||
299 | switch (err) { | |
300 | case -NFS4ERR_RESET_TO_MDS: | |
301 | filelayout_reset_write(data); | |
302 | return task->tk_status; | |
303 | case -EAGAIN: | |
d00c5d43 | 304 | rpc_restart_call_prepare(task); |
a69aef14 FI |
305 | return -EAGAIN; |
306 | } | |
307 | ||
863a3c6c | 308 | filelayout_set_layoutcommit(data); |
a69aef14 FI |
309 | return 0; |
310 | } | |
311 | ||
e0c2b380 | 312 | /* Fake up some data that will cause nfs_commit_release to retry the writes. */ |
0b7c0153 | 313 | static void prepare_to_resend_writes(struct nfs_commit_data *data) |
e0c2b380 FI |
314 | { |
315 | struct nfs_page *first = nfs_list_entry(data->pages.next); | |
316 | ||
317 | data->task.tk_status = 0; | |
318 | memcpy(data->verf.verifier, first->wb_verf.verifier, | |
319 | sizeof(first->wb_verf.verifier)); | |
320 | data->verf.verifier[0]++; /* ensure verifier mismatch */ | |
321 | } | |
322 | ||
323 | static int filelayout_commit_done_cb(struct rpc_task *task, | |
0b7c0153 | 324 | struct nfs_commit_data *data) |
e0c2b380 | 325 | { |
e7dd79af AA |
326 | int err; |
327 | ||
328 | err = filelayout_async_handle_error(task, NULL, data->ds_clp, | |
329 | data->lseg); | |
330 | ||
331 | switch (err) { | |
332 | case -NFS4ERR_RESET_TO_MDS: | |
333 | prepare_to_resend_writes(data); | |
334 | return -EAGAIN; | |
335 | case -EAGAIN: | |
336 | rpc_restart_call_prepare(task); | |
e0c2b380 FI |
337 | return -EAGAIN; |
338 | } | |
339 | ||
340 | return 0; | |
341 | } | |
342 | ||
a69aef14 FI |
343 | static void filelayout_write_prepare(struct rpc_task *task, void *data) |
344 | { | |
cd12ae32 | 345 | struct nfs_write_data *wdata = data; |
a69aef14 FI |
346 | |
347 | if (nfs41_setup_sequence(wdata->ds_clp->cl_session, | |
348 | &wdata->args.seq_args, &wdata->res.seq_res, | |
9d12b216 | 349 | task)) |
a69aef14 FI |
350 | return; |
351 | ||
352 | rpc_call_start(task); | |
353 | } | |
354 | ||
355 | static void filelayout_write_call_done(struct rpc_task *task, void *data) | |
356 | { | |
cd12ae32 | 357 | struct nfs_write_data *wdata = data; |
a69aef14 FI |
358 | |
359 | /* Note this may cause RPC to be resent */ | |
cd841605 | 360 | wdata->header->mds_ops->rpc_call_done(task, data); |
a69aef14 FI |
361 | } |
362 | ||
0a702195 WAA |
363 | static void filelayout_write_count_stats(struct rpc_task *task, void *data) |
364 | { | |
cd12ae32 | 365 | struct nfs_write_data *wdata = data; |
0a702195 | 366 | |
cd841605 | 367 | rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics); |
0a702195 WAA |
368 | } |
369 | ||
a69aef14 FI |
370 | static void filelayout_write_release(void *data) |
371 | { | |
cd12ae32 | 372 | struct nfs_write_data *wdata = data; |
a69aef14 | 373 | |
cd841605 | 374 | wdata->header->mds_ops->rpc_release(data); |
a69aef14 FI |
375 | } |
376 | ||
0b7c0153 | 377 | static void filelayout_commit_prepare(struct rpc_task *task, void *data) |
e0c2b380 | 378 | { |
0b7c0153 | 379 | struct nfs_commit_data *wdata = data; |
e0c2b380 | 380 | |
0b7c0153 FI |
381 | if (nfs41_setup_sequence(wdata->ds_clp->cl_session, |
382 | &wdata->args.seq_args, &wdata->res.seq_res, | |
383 | task)) | |
384 | return; | |
385 | ||
386 | rpc_call_start(task); | |
387 | } | |
388 | ||
389 | static void filelayout_write_commit_done(struct rpc_task *task, void *data) | |
390 | { | |
391 | struct nfs_commit_data *wdata = data; | |
392 | ||
393 | /* Note this may cause RPC to be resent */ | |
394 | wdata->mds_ops->rpc_call_done(task, data); | |
395 | } | |
396 | ||
397 | static void filelayout_commit_count_stats(struct rpc_task *task, void *data) | |
398 | { | |
399 | struct nfs_commit_data *cdata = data; | |
400 | ||
401 | rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics); | |
402 | } | |
403 | ||
404 | static void filelayout_commit_release(void *calldata) | |
405 | { | |
406 | struct nfs_commit_data *data = calldata; | |
407 | ||
f453a54a | 408 | data->completion_ops->completion(data); |
0b7c0153 FI |
409 | put_lseg(data->lseg); |
410 | nfs_commitdata_release(data); | |
e0c2b380 FI |
411 | } |
412 | ||
17280175 | 413 | static const struct rpc_call_ops filelayout_read_call_ops = { |
dc70d7b3 AA |
414 | .rpc_call_prepare = filelayout_read_prepare, |
415 | .rpc_call_done = filelayout_read_call_done, | |
0a702195 | 416 | .rpc_count_stats = filelayout_read_count_stats, |
dc70d7b3 AA |
417 | .rpc_release = filelayout_read_release, |
418 | }; | |
419 | ||
17280175 | 420 | static const struct rpc_call_ops filelayout_write_call_ops = { |
a69aef14 FI |
421 | .rpc_call_prepare = filelayout_write_prepare, |
422 | .rpc_call_done = filelayout_write_call_done, | |
0a702195 | 423 | .rpc_count_stats = filelayout_write_count_stats, |
a69aef14 FI |
424 | .rpc_release = filelayout_write_release, |
425 | }; | |
426 | ||
17280175 | 427 | static const struct rpc_call_ops filelayout_commit_call_ops = { |
0b7c0153 FI |
428 | .rpc_call_prepare = filelayout_commit_prepare, |
429 | .rpc_call_done = filelayout_write_commit_done, | |
430 | .rpc_count_stats = filelayout_commit_count_stats, | |
e0c2b380 FI |
431 | .rpc_release = filelayout_commit_release, |
432 | }; | |
433 | ||
dc70d7b3 AA |
434 | static enum pnfs_try_status |
435 | filelayout_read_pagelist(struct nfs_read_data *data) | |
436 | { | |
cd841605 FI |
437 | struct nfs_pgio_header *hdr = data->header; |
438 | struct pnfs_layout_segment *lseg = hdr->lseg; | |
dc70d7b3 AA |
439 | struct nfs4_pnfs_ds *ds; |
440 | loff_t offset = data->args.offset; | |
441 | u32 j, idx; | |
442 | struct nfs_fh *fh; | |
443 | int status; | |
444 | ||
445 | dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n", | |
cd841605 | 446 | __func__, hdr->inode->i_ino, |
dc70d7b3 AA |
447 | data->args.pgbase, (size_t)data->args.count, offset); |
448 | ||
449 | /* Retrieve the correct rpc_client for the byte range */ | |
450 | j = nfs4_fl_calc_j_index(lseg, offset); | |
451 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
452 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 453 | if (!ds) |
dc70d7b3 | 454 | return PNFS_NOT_ATTEMPTED; |
c9895cb6 | 455 | dprintk("%s USE DS: %s\n", __func__, ds->ds_remotestr); |
dc70d7b3 AA |
456 | |
457 | /* No multipath support. Use first DS */ | |
458 | data->ds_clp = ds->ds_clp; | |
459 | fh = nfs4_fl_select_ds_fh(lseg, j); | |
460 | if (fh) | |
461 | data->args.fh = fh; | |
462 | ||
463 | data->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
464 | data->mds_offset = offset; | |
465 | ||
466 | /* Perform an asynchronous read to ds */ | |
c5996c4e | 467 | status = nfs_initiate_read(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 | 468 | &filelayout_read_call_ops, RPC_TASK_SOFTCONN); |
dc70d7b3 AA |
469 | BUG_ON(status != 0); |
470 | return PNFS_ATTEMPTED; | |
471 | } | |
472 | ||
a69aef14 | 473 | /* Perform async writes. */ |
0382b744 AA |
474 | static enum pnfs_try_status |
475 | filelayout_write_pagelist(struct nfs_write_data *data, int sync) | |
476 | { | |
cd841605 FI |
477 | struct nfs_pgio_header *hdr = data->header; |
478 | struct pnfs_layout_segment *lseg = hdr->lseg; | |
a69aef14 FI |
479 | struct nfs4_pnfs_ds *ds; |
480 | loff_t offset = data->args.offset; | |
481 | u32 j, idx; | |
482 | struct nfs_fh *fh; | |
483 | int status; | |
484 | ||
485 | /* Retrieve the correct rpc_client for the byte range */ | |
486 | j = nfs4_fl_calc_j_index(lseg, offset); | |
487 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
488 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 489 | if (!ds) |
a69aef14 | 490 | return PNFS_NOT_ATTEMPTED; |
c9895cb6 | 491 | dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s\n", __func__, |
cd841605 | 492 | hdr->inode->i_ino, sync, (size_t) data->args.count, offset, |
c9895cb6 | 493 | ds->ds_remotestr); |
a69aef14 | 494 | |
a69aef14 FI |
495 | data->write_done_cb = filelayout_write_done_cb; |
496 | data->ds_clp = ds->ds_clp; | |
497 | fh = nfs4_fl_select_ds_fh(lseg, j); | |
498 | if (fh) | |
499 | data->args.fh = fh; | |
500 | /* | |
501 | * Get the file offset on the dserver. Set the write offset to | |
502 | * this offset and save the original offset. | |
503 | */ | |
504 | data->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
a69aef14 FI |
505 | |
506 | /* Perform an asynchronous write */ | |
c5996c4e | 507 | status = nfs_initiate_write(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 AA |
508 | &filelayout_write_call_ops, sync, |
509 | RPC_TASK_SOFTCONN); | |
a69aef14 FI |
510 | BUG_ON(status != 0); |
511 | return PNFS_ATTEMPTED; | |
0382b744 AA |
512 | } |
513 | ||
16b374ca AA |
514 | /* |
515 | * filelayout_check_layout() | |
516 | * | |
517 | * Make sure layout segment parameters are sane WRT the device. | |
518 | * At this point no generic layer initialization of the lseg has occurred, | |
519 | * and nothing has been added to the layout_hdr cache. | |
520 | * | |
521 | */ | |
522 | static int | |
523 | filelayout_check_layout(struct pnfs_layout_hdr *lo, | |
524 | struct nfs4_filelayout_segment *fl, | |
525 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 TM |
526 | struct nfs4_deviceid *id, |
527 | gfp_t gfp_flags) | |
16b374ca | 528 | { |
a1eaecbc | 529 | struct nfs4_deviceid_node *d; |
16b374ca AA |
530 | struct nfs4_file_layout_dsaddr *dsaddr; |
531 | int status = -EINVAL; | |
b7edfaa1 | 532 | struct nfs_server *nfss = NFS_SERVER(lo->plh_inode); |
16b374ca AA |
533 | |
534 | dprintk("--> %s\n", __func__); | |
535 | ||
7c24d948 AA |
536 | /* FIXME: remove this check when layout segment support is added */ |
537 | if (lgr->range.offset != 0 || | |
538 | lgr->range.length != NFS4_MAX_UINT64) { | |
539 | dprintk("%s Only whole file layouts supported. Use MDS i/o\n", | |
540 | __func__); | |
541 | goto out; | |
542 | } | |
543 | ||
16b374ca | 544 | if (fl->pattern_offset > lgr->range.offset) { |
3b6445a6 | 545 | dprintk("%s pattern_offset %lld too large\n", |
16b374ca AA |
546 | __func__, fl->pattern_offset); |
547 | goto out; | |
548 | } | |
549 | ||
75247aff BH |
550 | if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) { |
551 | dprintk("%s Invalid stripe unit (%u)\n", | |
16b374ca AA |
552 | __func__, fl->stripe_unit); |
553 | goto out; | |
554 | } | |
555 | ||
556 | /* find and reference the deviceid */ | |
35c8bb54 BH |
557 | d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld, |
558 | NFS_SERVER(lo->plh_inode)->nfs_client, id); | |
a1eaecbc | 559 | if (d == NULL) { |
a75b9df9 | 560 | dsaddr = get_device_info(lo->plh_inode, id, gfp_flags); |
16b374ca AA |
561 | if (dsaddr == NULL) |
562 | goto out; | |
a1eaecbc BH |
563 | } else |
564 | dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node); | |
c47abcf8 AA |
565 | /* Found deviceid is being reaped */ |
566 | if (test_bit(NFS_DEVICEID_INVALID, &dsaddr->id_node.flags)) | |
567 | goto out_put; | |
568 | ||
16b374ca AA |
569 | fl->dsaddr = dsaddr; |
570 | ||
e414966b CL |
571 | if (fl->first_stripe_index >= dsaddr->stripe_count) { |
572 | dprintk("%s Bad first_stripe_index %u\n", | |
16b374ca AA |
573 | __func__, fl->first_stripe_index); |
574 | goto out_put; | |
575 | } | |
576 | ||
577 | if ((fl->stripe_type == STRIPE_SPARSE && | |
578 | fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) || | |
579 | (fl->stripe_type == STRIPE_DENSE && | |
580 | fl->num_fh != dsaddr->stripe_count)) { | |
581 | dprintk("%s num_fh %u not valid for given packing\n", | |
582 | __func__, fl->num_fh); | |
583 | goto out_put; | |
584 | } | |
585 | ||
586 | if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) { | |
587 | dprintk("%s Stripe unit (%u) not aligned with rsize %u " | |
588 | "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize, | |
589 | nfss->wsize); | |
590 | } | |
591 | ||
592 | status = 0; | |
593 | out: | |
594 | dprintk("--> %s returns %d\n", __func__, status); | |
595 | return status; | |
596 | out_put: | |
ea8eecdd | 597 | nfs4_fl_put_deviceid(dsaddr); |
16b374ca AA |
598 | goto out; |
599 | } | |
600 | ||
601 | static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl) | |
602 | { | |
603 | int i; | |
604 | ||
605 | for (i = 0; i < fl->num_fh; i++) { | |
606 | if (!fl->fh_array[i]) | |
607 | break; | |
608 | kfree(fl->fh_array[i]); | |
609 | } | |
610 | kfree(fl->fh_array); | |
611 | fl->fh_array = NULL; | |
612 | } | |
613 | ||
614 | static void | |
615 | _filelayout_free_lseg(struct nfs4_filelayout_segment *fl) | |
616 | { | |
617 | filelayout_free_fh_array(fl); | |
618 | kfree(fl); | |
619 | } | |
620 | ||
621 | static int | |
622 | filelayout_decode_layout(struct pnfs_layout_hdr *flo, | |
623 | struct nfs4_filelayout_segment *fl, | |
624 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 TM |
625 | struct nfs4_deviceid *id, |
626 | gfp_t gfp_flags) | |
16b374ca | 627 | { |
35124a09 | 628 | struct xdr_stream stream; |
f7da7a12 | 629 | struct xdr_buf buf; |
35124a09 WAA |
630 | struct page *scratch; |
631 | __be32 *p; | |
16b374ca AA |
632 | uint32_t nfl_util; |
633 | int i; | |
634 | ||
635 | dprintk("%s: set_layout_map Begin\n", __func__); | |
636 | ||
a75b9df9 | 637 | scratch = alloc_page(gfp_flags); |
35124a09 WAA |
638 | if (!scratch) |
639 | return -ENOMEM; | |
640 | ||
f7da7a12 | 641 | xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len); |
35124a09 WAA |
642 | xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE); |
643 | ||
644 | /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8), | |
645 | * num_fh (4) */ | |
646 | p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20); | |
647 | if (unlikely(!p)) | |
648 | goto out_err; | |
649 | ||
16b374ca AA |
650 | memcpy(id, p, sizeof(*id)); |
651 | p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE); | |
a1eaecbc | 652 | nfs4_print_deviceid(id); |
16b374ca AA |
653 | |
654 | nfl_util = be32_to_cpup(p++); | |
655 | if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS) | |
656 | fl->commit_through_mds = 1; | |
657 | if (nfl_util & NFL4_UFLG_DENSE) | |
658 | fl->stripe_type = STRIPE_DENSE; | |
659 | else | |
660 | fl->stripe_type = STRIPE_SPARSE; | |
661 | fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK; | |
662 | ||
663 | fl->first_stripe_index = be32_to_cpup(p++); | |
664 | p = xdr_decode_hyper(p, &fl->pattern_offset); | |
665 | fl->num_fh = be32_to_cpup(p++); | |
666 | ||
667 | dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n", | |
668 | __func__, nfl_util, fl->num_fh, fl->first_stripe_index, | |
669 | fl->pattern_offset); | |
670 | ||
cec765cf AA |
671 | /* Note that a zero value for num_fh is legal for STRIPE_SPARSE. |
672 | * Futher checking is done in filelayout_check_layout */ | |
e414966b | 673 | if (fl->num_fh > |
cec765cf | 674 | max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT)) |
35124a09 WAA |
675 | goto out_err; |
676 | ||
cec765cf AA |
677 | if (fl->num_fh > 0) { |
678 | fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *), | |
679 | gfp_flags); | |
680 | if (!fl->fh_array) | |
681 | goto out_err; | |
682 | } | |
16b374ca AA |
683 | |
684 | for (i = 0; i < fl->num_fh; i++) { | |
685 | /* Do we want to use a mempool here? */ | |
a75b9df9 | 686 | fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags); |
35124a09 WAA |
687 | if (!fl->fh_array[i]) |
688 | goto out_err_free; | |
689 | ||
690 | p = xdr_inline_decode(&stream, 4); | |
691 | if (unlikely(!p)) | |
692 | goto out_err_free; | |
16b374ca AA |
693 | fl->fh_array[i]->size = be32_to_cpup(p++); |
694 | if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) { | |
f9fd2d9c | 695 | printk(KERN_ERR "NFS: Too big fh %d received %d\n", |
16b374ca | 696 | i, fl->fh_array[i]->size); |
35124a09 | 697 | goto out_err_free; |
16b374ca | 698 | } |
35124a09 WAA |
699 | |
700 | p = xdr_inline_decode(&stream, fl->fh_array[i]->size); | |
701 | if (unlikely(!p)) | |
702 | goto out_err_free; | |
16b374ca | 703 | memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size); |
16b374ca AA |
704 | dprintk("DEBUG: %s: fh len %d\n", __func__, |
705 | fl->fh_array[i]->size); | |
706 | } | |
707 | ||
35124a09 | 708 | __free_page(scratch); |
16b374ca | 709 | return 0; |
35124a09 WAA |
710 | |
711 | out_err_free: | |
712 | filelayout_free_fh_array(fl); | |
713 | out_err: | |
714 | __free_page(scratch); | |
715 | return -EIO; | |
16b374ca AA |
716 | } |
717 | ||
c879513e FI |
718 | static void |
719 | filelayout_free_lseg(struct pnfs_layout_segment *lseg) | |
720 | { | |
721 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
722 | ||
723 | dprintk("--> %s\n", __func__); | |
724 | nfs4_fl_put_deviceid(fl->dsaddr); | |
799ba8d5 FI |
725 | /* This assumes a single RW lseg */ |
726 | if (lseg->pls_range.iomode == IOMODE_RW) { | |
727 | struct nfs4_filelayout *flo; | |
728 | ||
729 | flo = FILELAYOUT_FROM_HDR(lseg->pls_layout); | |
730 | flo->commit_info.nbuckets = 0; | |
731 | kfree(flo->commit_info.buckets); | |
732 | flo->commit_info.buckets = NULL; | |
733 | } | |
c879513e FI |
734 | _filelayout_free_lseg(fl); |
735 | } | |
736 | ||
799ba8d5 FI |
737 | static int |
738 | filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg, | |
ea2cf228 | 739 | struct nfs_commit_info *cinfo, |
799ba8d5 FI |
740 | gfp_t gfp_flags) |
741 | { | |
742 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
ea2cf228 | 743 | struct pnfs_commit_bucket *buckets; |
799ba8d5 FI |
744 | int size; |
745 | ||
746 | if (fl->commit_through_mds) | |
747 | return 0; | |
ea2cf228 | 748 | if (cinfo->ds->nbuckets != 0) { |
799ba8d5 FI |
749 | /* This assumes there is only one IOMODE_RW lseg. What |
750 | * we really want to do is have a layout_hdr level | |
751 | * dictionary of <multipath_list4, fh> keys, each | |
752 | * associated with a struct list_head, populated by calls | |
753 | * to filelayout_write_pagelist(). | |
754 | * */ | |
755 | return 0; | |
756 | } | |
757 | ||
758 | size = (fl->stripe_type == STRIPE_SPARSE) ? | |
759 | fl->dsaddr->ds_num : fl->dsaddr->stripe_count; | |
760 | ||
ea2cf228 | 761 | buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket), |
799ba8d5 FI |
762 | gfp_flags); |
763 | if (!buckets) | |
764 | return -ENOMEM; | |
765 | else { | |
766 | int i; | |
767 | ||
ea2cf228 FI |
768 | spin_lock(cinfo->lock); |
769 | if (cinfo->ds->nbuckets != 0) | |
799ba8d5 FI |
770 | kfree(buckets); |
771 | else { | |
ea2cf228 FI |
772 | cinfo->ds->buckets = buckets; |
773 | cinfo->ds->nbuckets = size; | |
799ba8d5 FI |
774 | for (i = 0; i < size; i++) { |
775 | INIT_LIST_HEAD(&buckets[i].written); | |
776 | INIT_LIST_HEAD(&buckets[i].committing); | |
777 | } | |
778 | } | |
ea2cf228 | 779 | spin_unlock(cinfo->lock); |
799ba8d5 FI |
780 | return 0; |
781 | } | |
782 | } | |
783 | ||
16b374ca AA |
784 | static struct pnfs_layout_segment * |
785 | filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid, | |
a75b9df9 TM |
786 | struct nfs4_layoutget_res *lgr, |
787 | gfp_t gfp_flags) | |
16b374ca AA |
788 | { |
789 | struct nfs4_filelayout_segment *fl; | |
790 | int rc; | |
791 | struct nfs4_deviceid id; | |
792 | ||
793 | dprintk("--> %s\n", __func__); | |
a75b9df9 | 794 | fl = kzalloc(sizeof(*fl), gfp_flags); |
16b374ca AA |
795 | if (!fl) |
796 | return NULL; | |
797 | ||
a75b9df9 TM |
798 | rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags); |
799 | if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) { | |
16b374ca AA |
800 | _filelayout_free_lseg(fl); |
801 | return NULL; | |
802 | } | |
803 | return &fl->generic_hdr; | |
804 | } | |
805 | ||
94ad1c80 FI |
806 | /* |
807 | * filelayout_pg_test(). Called by nfs_can_coalesce_requests() | |
808 | * | |
18ad0a9f BH |
809 | * return true : coalesce page |
810 | * return false : don't coalesce page | |
94ad1c80 | 811 | */ |
1751c363 | 812 | static bool |
94ad1c80 FI |
813 | filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev, |
814 | struct nfs_page *req) | |
815 | { | |
816 | u64 p_stripe, r_stripe; | |
817 | u32 stripe_unit; | |
818 | ||
19345cb2 BH |
819 | if (!pnfs_generic_pg_test(pgio, prev, req) || |
820 | !nfs_generic_pg_test(pgio, prev, req)) | |
821 | return false; | |
89a58e32 | 822 | |
b5542849 FI |
823 | p_stripe = (u64)req_offset(prev); |
824 | r_stripe = (u64)req_offset(req); | |
94ad1c80 FI |
825 | stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit; |
826 | ||
827 | do_div(p_stripe, stripe_unit); | |
828 | do_div(r_stripe, stripe_unit); | |
829 | ||
830 | return (p_stripe == r_stripe); | |
831 | } | |
832 | ||
17280175 | 833 | static void |
7c24d948 AA |
834 | filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio, |
835 | struct nfs_page *req) | |
836 | { | |
837 | BUG_ON(pgio->pg_lseg != NULL); | |
838 | ||
1825a0d0 FI |
839 | if (req->wb_offset != req->wb_pgbase) { |
840 | /* | |
841 | * Handling unaligned pages is difficult, because have to | |
842 | * somehow split a req in two in certain cases in the | |
843 | * pg.test code. Avoid this by just not using pnfs | |
844 | * in this case. | |
845 | */ | |
846 | nfs_pageio_reset_read_mds(pgio); | |
847 | return; | |
848 | } | |
7c24d948 AA |
849 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
850 | req->wb_context, | |
851 | 0, | |
852 | NFS4_MAX_UINT64, | |
853 | IOMODE_READ, | |
854 | GFP_KERNEL); | |
855 | /* If no lseg, fall back to read through mds */ | |
856 | if (pgio->pg_lseg == NULL) | |
1f945357 | 857 | nfs_pageio_reset_read_mds(pgio); |
7c24d948 AA |
858 | } |
859 | ||
17280175 | 860 | static void |
7c24d948 AA |
861 | filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio, |
862 | struct nfs_page *req) | |
863 | { | |
ea2cf228 | 864 | struct nfs_commit_info cinfo; |
799ba8d5 FI |
865 | int status; |
866 | ||
7c24d948 AA |
867 | BUG_ON(pgio->pg_lseg != NULL); |
868 | ||
1825a0d0 FI |
869 | if (req->wb_offset != req->wb_pgbase) |
870 | goto out_mds; | |
7c24d948 AA |
871 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
872 | req->wb_context, | |
873 | 0, | |
874 | NFS4_MAX_UINT64, | |
875 | IOMODE_RW, | |
876 | GFP_NOFS); | |
877 | /* If no lseg, fall back to write through mds */ | |
878 | if (pgio->pg_lseg == NULL) | |
799ba8d5 | 879 | goto out_mds; |
ea2cf228 FI |
880 | nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq); |
881 | status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS); | |
799ba8d5 FI |
882 | if (status < 0) { |
883 | put_lseg(pgio->pg_lseg); | |
884 | pgio->pg_lseg = NULL; | |
885 | goto out_mds; | |
886 | } | |
887 | return; | |
888 | out_mds: | |
889 | nfs_pageio_reset_write_mds(pgio); | |
7c24d948 AA |
890 | } |
891 | ||
1751c363 | 892 | static const struct nfs_pageio_ops filelayout_pg_read_ops = { |
7c24d948 | 893 | .pg_init = filelayout_pg_init_read, |
1751c363 | 894 | .pg_test = filelayout_pg_test, |
493292dd | 895 | .pg_doio = pnfs_generic_pg_readpages, |
1751c363 TM |
896 | }; |
897 | ||
898 | static const struct nfs_pageio_ops filelayout_pg_write_ops = { | |
7c24d948 | 899 | .pg_init = filelayout_pg_init_write, |
1751c363 | 900 | .pg_test = filelayout_pg_test, |
dce81290 | 901 | .pg_doio = pnfs_generic_pg_writepages, |
1751c363 TM |
902 | }; |
903 | ||
e0c2b380 FI |
904 | static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j) |
905 | { | |
906 | if (fl->stripe_type == STRIPE_SPARSE) | |
907 | return nfs4_fl_calc_ds_index(&fl->generic_hdr, j); | |
908 | else | |
909 | return j; | |
910 | } | |
911 | ||
d6d6dc7c FI |
912 | /* The generic layer is about to remove the req from the commit list. |
913 | * If this will make the bucket empty, it will need to put the lseg reference. | |
d6d6dc7c | 914 | */ |
8dd37758 | 915 | static void |
ea2cf228 FI |
916 | filelayout_clear_request_commit(struct nfs_page *req, |
917 | struct nfs_commit_info *cinfo) | |
d6d6dc7c | 918 | { |
8dd37758 | 919 | struct pnfs_layout_segment *freeme = NULL; |
8dd37758 | 920 | |
ea2cf228 | 921 | spin_lock(cinfo->lock); |
8dd37758 TM |
922 | if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags)) |
923 | goto out; | |
ea2cf228 | 924 | cinfo->ds->nwritten--; |
d6d6dc7c | 925 | if (list_is_singular(&req->wb_list)) { |
ea2cf228 | 926 | struct pnfs_commit_bucket *bucket; |
d6d6dc7c | 927 | |
799ba8d5 | 928 | bucket = list_first_entry(&req->wb_list, |
ea2cf228 | 929 | struct pnfs_commit_bucket, |
799ba8d5 FI |
930 | written); |
931 | freeme = bucket->wlseg; | |
932 | bucket->wlseg = NULL; | |
d6d6dc7c | 933 | } |
8dd37758 | 934 | out: |
ea2cf228 FI |
935 | nfs_request_remove_commit_list(req, cinfo); |
936 | spin_unlock(cinfo->lock); | |
8dd37758 | 937 | put_lseg(freeme); |
d6d6dc7c FI |
938 | } |
939 | ||
940 | static struct list_head * | |
941 | filelayout_choose_commit_list(struct nfs_page *req, | |
ea2cf228 FI |
942 | struct pnfs_layout_segment *lseg, |
943 | struct nfs_commit_info *cinfo) | |
e0c2b380 | 944 | { |
e0c2b380 FI |
945 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); |
946 | u32 i, j; | |
947 | struct list_head *list; | |
ea2cf228 | 948 | struct pnfs_commit_bucket *buckets; |
e0c2b380 | 949 | |
d6d6dc7c | 950 | if (fl->commit_through_mds) |
ea2cf228 | 951 | return &cinfo->mds->list; |
d6d6dc7c | 952 | |
e0c2b380 FI |
953 | /* Note that we are calling nfs4_fl_calc_j_index on each page |
954 | * that ends up being committed to a data server. An attractive | |
955 | * alternative is to add a field to nfs_write_data and nfs_page | |
956 | * to store the value calculated in filelayout_write_pagelist | |
957 | * and just use that here. | |
958 | */ | |
b5542849 | 959 | j = nfs4_fl_calc_j_index(lseg, req_offset(req)); |
e0c2b380 | 960 | i = select_bucket_index(fl, j); |
ea2cf228 | 961 | buckets = cinfo->ds->buckets; |
799ba8d5 | 962 | list = &buckets[i].written; |
e0c2b380 | 963 | if (list_empty(list)) { |
d6d6dc7c FI |
964 | /* Non-empty buckets hold a reference on the lseg. That ref |
965 | * is normally transferred to the COMMIT call and released | |
966 | * there. It could also be released if the last req is pulled | |
967 | * off due to a rewrite, in which case it will be done in | |
799ba8d5 | 968 | * filelayout_clear_request_commit |
d6d6dc7c | 969 | */ |
799ba8d5 | 970 | buckets[i].wlseg = get_lseg(lseg); |
e0c2b380 | 971 | } |
8dd37758 | 972 | set_bit(PG_COMMIT_TO_DS, &req->wb_flags); |
ea2cf228 | 973 | cinfo->ds->nwritten++; |
e0c2b380 FI |
974 | return list; |
975 | } | |
976 | ||
8dd37758 TM |
977 | static void |
978 | filelayout_mark_request_commit(struct nfs_page *req, | |
ea2cf228 FI |
979 | struct pnfs_layout_segment *lseg, |
980 | struct nfs_commit_info *cinfo) | |
8dd37758 TM |
981 | { |
982 | struct list_head *list; | |
983 | ||
ea2cf228 FI |
984 | list = filelayout_choose_commit_list(req, lseg, cinfo); |
985 | nfs_request_add_commit_list(req, list, cinfo); | |
8dd37758 TM |
986 | } |
987 | ||
e0c2b380 FI |
988 | static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i) |
989 | { | |
990 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
991 | ||
992 | if (flseg->stripe_type == STRIPE_SPARSE) | |
993 | return i; | |
994 | else | |
995 | return nfs4_fl_calc_ds_index(lseg, i); | |
996 | } | |
997 | ||
998 | static struct nfs_fh * | |
999 | select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i) | |
1000 | { | |
1001 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1002 | ||
1003 | if (flseg->stripe_type == STRIPE_SPARSE) { | |
1004 | if (flseg->num_fh == 1) | |
1005 | i = 0; | |
1006 | else if (flseg->num_fh == 0) | |
1007 | /* Use the MDS OPEN fh set in nfs_read_rpcsetup */ | |
1008 | return NULL; | |
1009 | } | |
1010 | return flseg->fh_array[i]; | |
1011 | } | |
1012 | ||
0b7c0153 | 1013 | static int filelayout_initiate_commit(struct nfs_commit_data *data, int how) |
e0c2b380 FI |
1014 | { |
1015 | struct pnfs_layout_segment *lseg = data->lseg; | |
1016 | struct nfs4_pnfs_ds *ds; | |
1017 | u32 idx; | |
1018 | struct nfs_fh *fh; | |
1019 | ||
1020 | idx = calc_ds_index_from_commit(lseg, data->ds_commit_index); | |
1021 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
1022 | if (!ds) { | |
e0c2b380 | 1023 | prepare_to_resend_writes(data); |
8dd37758 | 1024 | filelayout_commit_release(data); |
e0c2b380 FI |
1025 | return -EAGAIN; |
1026 | } | |
1027 | dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how); | |
0b7c0153 | 1028 | data->commit_done_cb = filelayout_commit_done_cb; |
e0c2b380 FI |
1029 | data->ds_clp = ds->ds_clp; |
1030 | fh = select_ds_fh_from_commit(lseg, data->ds_commit_index); | |
1031 | if (fh) | |
1032 | data->args.fh = fh; | |
0b7c0153 | 1033 | return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 AA |
1034 | &filelayout_commit_call_ops, how, |
1035 | RPC_TASK_SOFTCONN); | |
e0c2b380 FI |
1036 | } |
1037 | ||
8dd37758 | 1038 | static int |
1763da12 FI |
1039 | transfer_commit_list(struct list_head *src, struct list_head *dst, |
1040 | struct nfs_commit_info *cinfo, int max) | |
8dd37758 | 1041 | { |
8dd37758 TM |
1042 | struct nfs_page *req, *tmp; |
1043 | int ret = 0; | |
1044 | ||
1045 | list_for_each_entry_safe(req, tmp, src, wb_list) { | |
1046 | if (!nfs_lock_request(req)) | |
1047 | continue; | |
ea2cf228 | 1048 | if (cond_resched_lock(cinfo->lock)) |
3b3be88d | 1049 | list_safe_reset_next(req, tmp, wb_list); |
ea2cf228 | 1050 | nfs_request_remove_commit_list(req, cinfo); |
8dd37758 TM |
1051 | clear_bit(PG_COMMIT_TO_DS, &req->wb_flags); |
1052 | nfs_list_add_request(req, dst); | |
1053 | ret++; | |
1763da12 | 1054 | if ((ret == max) && !cinfo->dreq) |
8dd37758 TM |
1055 | break; |
1056 | } | |
1763da12 FI |
1057 | return ret; |
1058 | } | |
1059 | ||
1060 | static int | |
1061 | filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket, | |
1062 | struct nfs_commit_info *cinfo, | |
1063 | int max) | |
1064 | { | |
1065 | struct list_head *src = &bucket->written; | |
1066 | struct list_head *dst = &bucket->committing; | |
1067 | int ret; | |
1068 | ||
1069 | ret = transfer_commit_list(src, dst, cinfo, max); | |
799ba8d5 | 1070 | if (ret) { |
ea2cf228 FI |
1071 | cinfo->ds->nwritten -= ret; |
1072 | cinfo->ds->ncommitting += ret; | |
799ba8d5 FI |
1073 | bucket->clseg = bucket->wlseg; |
1074 | if (list_empty(src)) | |
1075 | bucket->wlseg = NULL; | |
1076 | else | |
1077 | get_lseg(bucket->clseg); | |
1078 | } | |
8dd37758 TM |
1079 | return ret; |
1080 | } | |
1081 | ||
d6d6dc7c | 1082 | /* Move reqs from written to committing lists, returning count of number moved. |
ea2cf228 | 1083 | * Note called with cinfo->lock held. |
d6d6dc7c | 1084 | */ |
ea2cf228 FI |
1085 | static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo, |
1086 | int max) | |
d6d6dc7c | 1087 | { |
d6d6dc7c FI |
1088 | int i, rv = 0, cnt; |
1089 | ||
ea2cf228 FI |
1090 | for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) { |
1091 | cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i], | |
1092 | cinfo, max); | |
d6d6dc7c FI |
1093 | max -= cnt; |
1094 | rv += cnt; | |
1095 | } | |
d6d6dc7c FI |
1096 | return rv; |
1097 | } | |
1098 | ||
1763da12 FI |
1099 | /* Pull everything off the committing lists and dump into @dst */ |
1100 | static void filelayout_recover_commit_reqs(struct list_head *dst, | |
1101 | struct nfs_commit_info *cinfo) | |
1102 | { | |
1103 | struct pnfs_commit_bucket *b; | |
1104 | int i; | |
1105 | ||
1106 | /* NOTE cinfo->lock is NOT held, relying on fact that this is | |
1107 | * only called on single thread per dreq. | |
1108 | * Can't take the lock because need to do put_lseg | |
1109 | */ | |
1110 | for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) { | |
1111 | if (transfer_commit_list(&b->written, dst, cinfo, 0)) { | |
1112 | BUG_ON(!list_empty(&b->written)); | |
1113 | put_lseg(b->wlseg); | |
1114 | b->wlseg = NULL; | |
1115 | } | |
1116 | } | |
1117 | cinfo->ds->nwritten = 0; | |
1118 | } | |
1119 | ||
9390f425 | 1120 | static unsigned int |
ea2cf228 | 1121 | alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list) |
e0c2b380 | 1122 | { |
ea2cf228 FI |
1123 | struct pnfs_ds_commit_info *fl_cinfo; |
1124 | struct pnfs_commit_bucket *bucket; | |
0b7c0153 | 1125 | struct nfs_commit_data *data; |
e0c2b380 | 1126 | int i, j; |
9390f425 | 1127 | unsigned int nreq = 0; |
e0c2b380 | 1128 | |
ea2cf228 | 1129 | fl_cinfo = cinfo->ds; |
799ba8d5 FI |
1130 | bucket = fl_cinfo->buckets; |
1131 | for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) { | |
1132 | if (list_empty(&bucket->committing)) | |
e0c2b380 FI |
1133 | continue; |
1134 | data = nfs_commitdata_alloc(); | |
1135 | if (!data) | |
9390f425 | 1136 | break; |
e0c2b380 | 1137 | data->ds_commit_index = i; |
799ba8d5 FI |
1138 | data->lseg = bucket->clseg; |
1139 | bucket->clseg = NULL; | |
e0c2b380 | 1140 | list_add(&data->pages, list); |
9390f425 | 1141 | nreq++; |
e0c2b380 | 1142 | } |
e0c2b380 | 1143 | |
9390f425 | 1144 | /* Clean up on error */ |
799ba8d5 FI |
1145 | for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) { |
1146 | if (list_empty(&bucket->committing)) | |
e0c2b380 | 1147 | continue; |
ea2cf228 | 1148 | nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo); |
799ba8d5 FI |
1149 | put_lseg(bucket->clseg); |
1150 | bucket->clseg = NULL; | |
e0c2b380 | 1151 | } |
e0c2b380 | 1152 | /* Caller will clean up entries put on list */ |
9390f425 | 1153 | return nreq; |
e0c2b380 FI |
1154 | } |
1155 | ||
1156 | /* This follows nfs_commit_list pretty closely */ | |
1157 | static int | |
1158 | filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages, | |
ea2cf228 | 1159 | int how, struct nfs_commit_info *cinfo) |
e0c2b380 | 1160 | { |
0b7c0153 | 1161 | struct nfs_commit_data *data, *tmp; |
e0c2b380 | 1162 | LIST_HEAD(list); |
9390f425 | 1163 | unsigned int nreq = 0; |
e0c2b380 FI |
1164 | |
1165 | if (!list_empty(mds_pages)) { | |
1166 | data = nfs_commitdata_alloc(); | |
9390f425 TM |
1167 | if (data != NULL) { |
1168 | data->lseg = NULL; | |
1169 | list_add(&data->pages, &list); | |
1170 | nreq++; | |
1171 | } else | |
ea2cf228 | 1172 | nfs_retry_commit(mds_pages, NULL, cinfo); |
e0c2b380 FI |
1173 | } |
1174 | ||
ea2cf228 | 1175 | nreq += alloc_ds_commits(cinfo, &list); |
9390f425 TM |
1176 | |
1177 | if (nreq == 0) { | |
f453a54a | 1178 | cinfo->completion_ops->error_cleanup(NFS_I(inode)); |
9390f425 TM |
1179 | goto out; |
1180 | } | |
1181 | ||
ea2cf228 | 1182 | atomic_add(nreq, &cinfo->mds->rpcs_out); |
e0c2b380 FI |
1183 | |
1184 | list_for_each_entry_safe(data, tmp, &list, pages) { | |
1185 | list_del_init(&data->pages); | |
e0c2b380 | 1186 | if (!data->lseg) { |
f453a54a | 1187 | nfs_init_commit(data, mds_pages, NULL, cinfo); |
0b7c0153 | 1188 | nfs_initiate_commit(NFS_CLIENT(inode), data, |
9f0ec176 | 1189 | data->mds_ops, how, 0); |
e0c2b380 | 1190 | } else { |
ea2cf228 | 1191 | struct pnfs_commit_bucket *buckets; |
799ba8d5 | 1192 | |
ea2cf228 | 1193 | buckets = cinfo->ds->buckets; |
f453a54a | 1194 | nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo); |
e0c2b380 FI |
1195 | filelayout_initiate_commit(data, how); |
1196 | } | |
1197 | } | |
9390f425 | 1198 | out: |
ea2cf228 | 1199 | cinfo->ds->ncommitting = 0; |
9390f425 | 1200 | return PNFS_ATTEMPTED; |
e0c2b380 FI |
1201 | } |
1202 | ||
1775bc34 BH |
1203 | static void |
1204 | filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d) | |
1205 | { | |
1206 | nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node)); | |
1207 | } | |
1208 | ||
799ba8d5 FI |
1209 | static struct pnfs_layout_hdr * |
1210 | filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags) | |
1211 | { | |
1212 | struct nfs4_filelayout *flo; | |
1213 | ||
1214 | flo = kzalloc(sizeof(*flo), gfp_flags); | |
1215 | return &flo->generic_hdr; | |
1216 | } | |
1217 | ||
1218 | static void | |
1219 | filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo) | |
1220 | { | |
1221 | kfree(FILELAYOUT_FROM_HDR(lo)); | |
1222 | } | |
1223 | ||
ea2cf228 FI |
1224 | static struct pnfs_ds_commit_info * |
1225 | filelayout_get_ds_info(struct inode *inode) | |
1226 | { | |
df011748 FI |
1227 | struct pnfs_layout_hdr *layout = NFS_I(inode)->layout; |
1228 | ||
1229 | if (layout == NULL) | |
1230 | return NULL; | |
1231 | else | |
1232 | return &FILELAYOUT_FROM_HDR(layout)->commit_info; | |
ea2cf228 FI |
1233 | } |
1234 | ||
7ab672ce | 1235 | static struct pnfs_layoutdriver_type filelayout_type = { |
ea8eecdd CH |
1236 | .id = LAYOUT_NFSV4_1_FILES, |
1237 | .name = "LAYOUT_NFSV4_1_FILES", | |
1238 | .owner = THIS_MODULE, | |
799ba8d5 FI |
1239 | .alloc_layout_hdr = filelayout_alloc_layout_hdr, |
1240 | .free_layout_hdr = filelayout_free_layout_hdr, | |
ea8eecdd CH |
1241 | .alloc_lseg = filelayout_alloc_lseg, |
1242 | .free_lseg = filelayout_free_lseg, | |
1751c363 TM |
1243 | .pg_read_ops = &filelayout_pg_read_ops, |
1244 | .pg_write_ops = &filelayout_pg_write_ops, | |
ea2cf228 | 1245 | .get_ds_info = &filelayout_get_ds_info, |
8dd37758 TM |
1246 | .mark_request_commit = filelayout_mark_request_commit, |
1247 | .clear_request_commit = filelayout_clear_request_commit, | |
d6d6dc7c | 1248 | .scan_commit_lists = filelayout_scan_commit_lists, |
1763da12 | 1249 | .recover_commit_reqs = filelayout_recover_commit_reqs, |
e0c2b380 | 1250 | .commit_pagelist = filelayout_commit_pagelist, |
dc70d7b3 | 1251 | .read_pagelist = filelayout_read_pagelist, |
0382b744 | 1252 | .write_pagelist = filelayout_write_pagelist, |
1775bc34 | 1253 | .free_deviceid_node = filelayout_free_deveiceid_node, |
7ab672ce DH |
1254 | }; |
1255 | ||
1256 | static int __init nfs4filelayout_init(void) | |
1257 | { | |
1258 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n", | |
1259 | __func__); | |
1260 | return pnfs_register_layoutdriver(&filelayout_type); | |
1261 | } | |
1262 | ||
1263 | static void __exit nfs4filelayout_exit(void) | |
1264 | { | |
1265 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n", | |
1266 | __func__); | |
1267 | pnfs_unregister_layoutdriver(&filelayout_type); | |
1268 | } | |
1269 | ||
f85ef69c BF |
1270 | MODULE_ALIAS("nfs-layouttype4-1"); |
1271 | ||
7ab672ce DH |
1272 | module_init(nfs4filelayout_init); |
1273 | module_exit(nfs4filelayout_exit); |