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 | } | |
3a7936c3 AA |
104 | /* balance nfs_get_client in filelayout_write_pagelist */ |
105 | nfs_put_client(data->ds_clp); | |
106 | data->ds_clp = NULL; | |
e7dd79af AA |
107 | } |
108 | ||
109 | static void filelayout_reset_read(struct nfs_read_data *data) | |
110 | { | |
111 | struct nfs_pgio_header *hdr = data->header; | |
112 | struct inode *inode = hdr->inode; | |
113 | struct rpc_task *task = &data->task; | |
114 | ||
115 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { | |
116 | dprintk("%s Reset task %5u for i/o through MDS " | |
117 | "(req %s/%lld, %u bytes @ offset %llu)\n", __func__, | |
118 | data->task.tk_pid, | |
119 | inode->i_sb->s_id, | |
120 | (long long)NFS_FILEID(inode), | |
121 | data->args.count, | |
122 | (unsigned long long)data->args.offset); | |
123 | ||
124 | task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode, | |
125 | &hdr->pages, | |
126 | hdr->completion_ops); | |
127 | } | |
3a7936c3 AA |
128 | /* balance nfs_get_client in filelayout_read_pagelist */ |
129 | nfs_put_client(data->ds_clp); | |
130 | data->ds_clp = NULL; | |
e7dd79af AA |
131 | } |
132 | ||
cbdabc7f AA |
133 | static int filelayout_async_handle_error(struct rpc_task *task, |
134 | struct nfs4_state *state, | |
135 | struct nfs_client *clp, | |
e7dd79af | 136 | struct pnfs_layout_segment *lseg) |
cbdabc7f | 137 | { |
e7dd79af AA |
138 | struct inode *inode = lseg->pls_layout->plh_inode; |
139 | struct nfs_server *mds_server = NFS_SERVER(inode); | |
140 | struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg); | |
9cb81968 | 141 | struct nfs_client *mds_client = mds_server->nfs_client; |
671fb896 | 142 | struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table; |
9cb81968 | 143 | |
cbdabc7f AA |
144 | if (task->tk_status >= 0) |
145 | return 0; | |
cbdabc7f AA |
146 | |
147 | switch (task->tk_status) { | |
9cb81968 AA |
148 | /* MDS state errors */ |
149 | case -NFS4ERR_DELEG_REVOKED: | |
150 | case -NFS4ERR_ADMIN_REVOKED: | |
151 | case -NFS4ERR_BAD_STATEID: | |
e7dd79af AA |
152 | if (state == NULL) |
153 | break; | |
2dc31756 | 154 | nfs_remove_bad_delegation(state->inode); |
9cb81968 | 155 | case -NFS4ERR_OPENMODE: |
e7dd79af AA |
156 | if (state == NULL) |
157 | break; | |
9cb81968 AA |
158 | nfs4_schedule_stateid_recovery(mds_server, state); |
159 | goto wait_on_recovery; | |
160 | case -NFS4ERR_EXPIRED: | |
e7dd79af AA |
161 | if (state != NULL) |
162 | nfs4_schedule_stateid_recovery(mds_server, state); | |
9cb81968 AA |
163 | nfs4_schedule_lease_recovery(mds_client); |
164 | goto wait_on_recovery; | |
165 | /* DS session errors */ | |
cbdabc7f AA |
166 | case -NFS4ERR_BADSESSION: |
167 | case -NFS4ERR_BADSLOT: | |
168 | case -NFS4ERR_BAD_HIGH_SLOT: | |
169 | case -NFS4ERR_DEADSESSION: | |
170 | case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION: | |
171 | case -NFS4ERR_SEQ_FALSE_RETRY: | |
172 | case -NFS4ERR_SEQ_MISORDERED: | |
173 | dprintk("%s ERROR %d, Reset session. Exchangeid " | |
174 | "flags 0x%x\n", __func__, task->tk_status, | |
175 | clp->cl_exchange_flags); | |
176 | nfs4_schedule_session_recovery(clp->cl_session); | |
177 | break; | |
178 | case -NFS4ERR_DELAY: | |
179 | case -NFS4ERR_GRACE: | |
180 | case -EKEYEXPIRED: | |
181 | rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX); | |
182 | break; | |
a8a4ae3a AA |
183 | case -NFS4ERR_RETRY_UNCACHED_REP: |
184 | break; | |
041245c8 AA |
185 | /* Invalidate Layout errors */ |
186 | case -NFS4ERR_PNFS_NO_LAYOUT: | |
187 | case -ESTALE: /* mapped NFS4ERR_STALE */ | |
188 | case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */ | |
189 | case -EISDIR: /* mapped NFS4ERR_ISDIR */ | |
190 | case -NFS4ERR_FHEXPIRED: | |
191 | case -NFS4ERR_WRONG_TYPE: | |
192 | dprintk("%s Invalid layout error %d\n", __func__, | |
193 | task->tk_status); | |
194 | /* | |
195 | * Destroy layout so new i/o will get a new layout. | |
196 | * Layout will not be destroyed until all current lseg | |
197 | * references are put. Mark layout as invalid to resend failed | |
198 | * i/o and all i/o waiting on the slot table to the MDS until | |
199 | * layout is destroyed and a new valid layout is obtained. | |
200 | */ | |
201 | set_bit(NFS_LAYOUT_INVALID, | |
202 | &NFS_I(state->inode)->layout->plh_flags); | |
203 | pnfs_destroy_layout(NFS_I(state->inode)); | |
204 | rpc_wake_up(&tbl->slot_tbl_waitq); | |
205 | goto reset; | |
e7dd79af AA |
206 | /* RPC connection errors */ |
207 | case -ECONNREFUSED: | |
208 | case -EHOSTDOWN: | |
209 | case -EHOSTUNREACH: | |
210 | case -ENETUNREACH: | |
211 | case -EIO: | |
212 | case -ETIMEDOUT: | |
213 | case -EPIPE: | |
214 | dprintk("%s DS connection error %d\n", __func__, | |
215 | task->tk_status); | |
0a57cdac AA |
216 | if (!filelayout_test_devid_invalid(devid)) |
217 | _pnfs_return_layout(state->inode); | |
e7dd79af | 218 | filelayout_mark_devid_invalid(devid); |
671fb896 | 219 | rpc_wake_up(&tbl->slot_tbl_waitq); |
b4a2967e | 220 | nfs4_ds_disconnect(clp); |
e7dd79af | 221 | /* fall through */ |
cbdabc7f | 222 | default: |
041245c8 | 223 | reset: |
e7dd79af | 224 | dprintk("%s Retry through MDS. Error %d\n", __func__, |
cbdabc7f | 225 | task->tk_status); |
e7dd79af | 226 | return -NFS4ERR_RESET_TO_MDS; |
cbdabc7f | 227 | } |
9cb81968 | 228 | out: |
cbdabc7f AA |
229 | task->tk_status = 0; |
230 | return -EAGAIN; | |
9cb81968 AA |
231 | wait_on_recovery: |
232 | rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL); | |
233 | if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0) | |
234 | rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task); | |
235 | goto out; | |
cbdabc7f AA |
236 | } |
237 | ||
238 | /* NFS_PROTO call done callback routines */ | |
239 | ||
240 | static int filelayout_read_done_cb(struct rpc_task *task, | |
241 | struct nfs_read_data *data) | |
242 | { | |
e7dd79af AA |
243 | struct nfs_pgio_header *hdr = data->header; |
244 | int err; | |
cbdabc7f | 245 | |
e7dd79af AA |
246 | err = filelayout_async_handle_error(task, data->args.context->state, |
247 | data->ds_clp, hdr->lseg); | |
cbdabc7f | 248 | |
e7dd79af AA |
249 | switch (err) { |
250 | case -NFS4ERR_RESET_TO_MDS: | |
251 | filelayout_reset_read(data); | |
252 | return task->tk_status; | |
253 | case -EAGAIN: | |
d00c5d43 | 254 | rpc_restart_call_prepare(task); |
cbdabc7f AA |
255 | return -EAGAIN; |
256 | } | |
257 | ||
258 | return 0; | |
259 | } | |
260 | ||
863a3c6c AA |
261 | /* |
262 | * We reference the rpc_cred of the first WRITE that triggers the need for | |
263 | * a LAYOUTCOMMIT, and use it to send the layoutcommit compound. | |
264 | * rfc5661 is not clear about which credential should be used. | |
265 | */ | |
266 | static void | |
267 | filelayout_set_layoutcommit(struct nfs_write_data *wdata) | |
268 | { | |
cd841605 FI |
269 | struct nfs_pgio_header *hdr = wdata->header; |
270 | ||
271 | if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds || | |
863a3c6c AA |
272 | wdata->res.verf->committed == NFS_FILE_SYNC) |
273 | return; | |
274 | ||
275 | pnfs_set_layoutcommit(wdata); | |
cd841605 FI |
276 | dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino, |
277 | (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb); | |
863a3c6c AA |
278 | } |
279 | ||
dc70d7b3 AA |
280 | /* |
281 | * Call ops for the async read/write cases | |
282 | * In the case of dense layouts, the offset needs to be reset to its | |
283 | * original value. | |
284 | */ | |
285 | static void filelayout_read_prepare(struct rpc_task *task, void *data) | |
286 | { | |
cd12ae32 | 287 | struct nfs_read_data *rdata = data; |
dc70d7b3 | 288 | |
041245c8 | 289 | if (filelayout_reset_to_mds(rdata->header->lseg)) { |
0ad2f378 AA |
290 | dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid); |
291 | filelayout_reset_read(rdata); | |
292 | rpc_exit(task, 0); | |
293 | return; | |
294 | } | |
cbdabc7f AA |
295 | rdata->read_done_cb = filelayout_read_done_cb; |
296 | ||
dc70d7b3 AA |
297 | if (nfs41_setup_sequence(rdata->ds_clp->cl_session, |
298 | &rdata->args.seq_args, &rdata->res.seq_res, | |
9d12b216 | 299 | task)) |
dc70d7b3 AA |
300 | return; |
301 | ||
302 | rpc_call_start(task); | |
303 | } | |
304 | ||
305 | static void filelayout_read_call_done(struct rpc_task *task, void *data) | |
306 | { | |
cd12ae32 | 307 | struct nfs_read_data *rdata = data; |
dc70d7b3 AA |
308 | |
309 | dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status); | |
310 | ||
0ad2f378 AA |
311 | if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags)) |
312 | return; | |
313 | ||
dc70d7b3 | 314 | /* Note this may cause RPC to be resent */ |
cd841605 | 315 | rdata->header->mds_ops->rpc_call_done(task, data); |
dc70d7b3 AA |
316 | } |
317 | ||
0a702195 WAA |
318 | static void filelayout_read_count_stats(struct rpc_task *task, void *data) |
319 | { | |
cd12ae32 | 320 | struct nfs_read_data *rdata = data; |
0a702195 | 321 | |
cd841605 | 322 | rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics); |
0a702195 WAA |
323 | } |
324 | ||
dc70d7b3 AA |
325 | static void filelayout_read_release(void *data) |
326 | { | |
cd12ae32 | 327 | struct nfs_read_data *rdata = data; |
dc70d7b3 | 328 | |
3a7936c3 AA |
329 | if (!test_bit(NFS_IOHDR_REDO, &rdata->header->flags)) |
330 | nfs_put_client(rdata->ds_clp); | |
cd841605 | 331 | rdata->header->mds_ops->rpc_release(data); |
dc70d7b3 AA |
332 | } |
333 | ||
a69aef14 FI |
334 | static int filelayout_write_done_cb(struct rpc_task *task, |
335 | struct nfs_write_data *data) | |
336 | { | |
e7dd79af AA |
337 | struct nfs_pgio_header *hdr = data->header; |
338 | int err; | |
339 | ||
340 | err = filelayout_async_handle_error(task, data->args.context->state, | |
341 | data->ds_clp, hdr->lseg); | |
342 | ||
343 | switch (err) { | |
344 | case -NFS4ERR_RESET_TO_MDS: | |
345 | filelayout_reset_write(data); | |
346 | return task->tk_status; | |
347 | case -EAGAIN: | |
d00c5d43 | 348 | rpc_restart_call_prepare(task); |
a69aef14 FI |
349 | return -EAGAIN; |
350 | } | |
351 | ||
863a3c6c | 352 | filelayout_set_layoutcommit(data); |
a69aef14 FI |
353 | return 0; |
354 | } | |
355 | ||
e0c2b380 | 356 | /* Fake up some data that will cause nfs_commit_release to retry the writes. */ |
0b7c0153 | 357 | static void prepare_to_resend_writes(struct nfs_commit_data *data) |
e0c2b380 FI |
358 | { |
359 | struct nfs_page *first = nfs_list_entry(data->pages.next); | |
360 | ||
361 | data->task.tk_status = 0; | |
362 | memcpy(data->verf.verifier, first->wb_verf.verifier, | |
363 | sizeof(first->wb_verf.verifier)); | |
364 | data->verf.verifier[0]++; /* ensure verifier mismatch */ | |
365 | } | |
366 | ||
367 | static int filelayout_commit_done_cb(struct rpc_task *task, | |
0b7c0153 | 368 | struct nfs_commit_data *data) |
e0c2b380 | 369 | { |
e7dd79af AA |
370 | int err; |
371 | ||
372 | err = filelayout_async_handle_error(task, NULL, data->ds_clp, | |
373 | data->lseg); | |
374 | ||
375 | switch (err) { | |
376 | case -NFS4ERR_RESET_TO_MDS: | |
377 | prepare_to_resend_writes(data); | |
378 | return -EAGAIN; | |
379 | case -EAGAIN: | |
380 | rpc_restart_call_prepare(task); | |
e0c2b380 FI |
381 | return -EAGAIN; |
382 | } | |
383 | ||
384 | return 0; | |
385 | } | |
386 | ||
a69aef14 FI |
387 | static void filelayout_write_prepare(struct rpc_task *task, void *data) |
388 | { | |
cd12ae32 | 389 | struct nfs_write_data *wdata = data; |
a69aef14 | 390 | |
041245c8 | 391 | if (filelayout_reset_to_mds(wdata->header->lseg)) { |
0ad2f378 AA |
392 | dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid); |
393 | filelayout_reset_write(wdata); | |
394 | rpc_exit(task, 0); | |
395 | return; | |
396 | } | |
a69aef14 FI |
397 | if (nfs41_setup_sequence(wdata->ds_clp->cl_session, |
398 | &wdata->args.seq_args, &wdata->res.seq_res, | |
9d12b216 | 399 | task)) |
a69aef14 FI |
400 | return; |
401 | ||
402 | rpc_call_start(task); | |
403 | } | |
404 | ||
405 | static void filelayout_write_call_done(struct rpc_task *task, void *data) | |
406 | { | |
cd12ae32 | 407 | struct nfs_write_data *wdata = data; |
a69aef14 | 408 | |
0ad2f378 AA |
409 | if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags)) |
410 | return; | |
411 | ||
a69aef14 | 412 | /* Note this may cause RPC to be resent */ |
cd841605 | 413 | wdata->header->mds_ops->rpc_call_done(task, data); |
a69aef14 FI |
414 | } |
415 | ||
0a702195 WAA |
416 | static void filelayout_write_count_stats(struct rpc_task *task, void *data) |
417 | { | |
cd12ae32 | 418 | struct nfs_write_data *wdata = data; |
0a702195 | 419 | |
cd841605 | 420 | rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics); |
0a702195 WAA |
421 | } |
422 | ||
a69aef14 FI |
423 | static void filelayout_write_release(void *data) |
424 | { | |
cd12ae32 | 425 | struct nfs_write_data *wdata = data; |
a69aef14 | 426 | |
3a7936c3 AA |
427 | if (!test_bit(NFS_IOHDR_REDO, &wdata->header->flags)) |
428 | nfs_put_client(wdata->ds_clp); | |
cd841605 | 429 | wdata->header->mds_ops->rpc_release(data); |
a69aef14 FI |
430 | } |
431 | ||
0b7c0153 | 432 | static void filelayout_commit_prepare(struct rpc_task *task, void *data) |
e0c2b380 | 433 | { |
0b7c0153 | 434 | struct nfs_commit_data *wdata = data; |
e0c2b380 | 435 | |
0b7c0153 FI |
436 | if (nfs41_setup_sequence(wdata->ds_clp->cl_session, |
437 | &wdata->args.seq_args, &wdata->res.seq_res, | |
438 | task)) | |
439 | return; | |
440 | ||
441 | rpc_call_start(task); | |
442 | } | |
443 | ||
444 | static void filelayout_write_commit_done(struct rpc_task *task, void *data) | |
445 | { | |
446 | struct nfs_commit_data *wdata = data; | |
447 | ||
448 | /* Note this may cause RPC to be resent */ | |
449 | wdata->mds_ops->rpc_call_done(task, data); | |
450 | } | |
451 | ||
452 | static void filelayout_commit_count_stats(struct rpc_task *task, void *data) | |
453 | { | |
454 | struct nfs_commit_data *cdata = data; | |
455 | ||
456 | rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics); | |
457 | } | |
458 | ||
459 | static void filelayout_commit_release(void *calldata) | |
460 | { | |
461 | struct nfs_commit_data *data = calldata; | |
462 | ||
f453a54a | 463 | data->completion_ops->completion(data); |
0b7c0153 | 464 | put_lseg(data->lseg); |
3a7936c3 | 465 | nfs_put_client(data->ds_clp); |
0b7c0153 | 466 | nfs_commitdata_release(data); |
e0c2b380 FI |
467 | } |
468 | ||
17280175 | 469 | static const struct rpc_call_ops filelayout_read_call_ops = { |
dc70d7b3 AA |
470 | .rpc_call_prepare = filelayout_read_prepare, |
471 | .rpc_call_done = filelayout_read_call_done, | |
0a702195 | 472 | .rpc_count_stats = filelayout_read_count_stats, |
dc70d7b3 AA |
473 | .rpc_release = filelayout_read_release, |
474 | }; | |
475 | ||
17280175 | 476 | static const struct rpc_call_ops filelayout_write_call_ops = { |
a69aef14 FI |
477 | .rpc_call_prepare = filelayout_write_prepare, |
478 | .rpc_call_done = filelayout_write_call_done, | |
0a702195 | 479 | .rpc_count_stats = filelayout_write_count_stats, |
a69aef14 FI |
480 | .rpc_release = filelayout_write_release, |
481 | }; | |
482 | ||
17280175 | 483 | static const struct rpc_call_ops filelayout_commit_call_ops = { |
0b7c0153 FI |
484 | .rpc_call_prepare = filelayout_commit_prepare, |
485 | .rpc_call_done = filelayout_write_commit_done, | |
486 | .rpc_count_stats = filelayout_commit_count_stats, | |
e0c2b380 FI |
487 | .rpc_release = filelayout_commit_release, |
488 | }; | |
489 | ||
dc70d7b3 AA |
490 | static enum pnfs_try_status |
491 | filelayout_read_pagelist(struct nfs_read_data *data) | |
492 | { | |
cd841605 FI |
493 | struct nfs_pgio_header *hdr = data->header; |
494 | struct pnfs_layout_segment *lseg = hdr->lseg; | |
dc70d7b3 AA |
495 | struct nfs4_pnfs_ds *ds; |
496 | loff_t offset = data->args.offset; | |
497 | u32 j, idx; | |
498 | struct nfs_fh *fh; | |
499 | int status; | |
500 | ||
501 | dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n", | |
cd841605 | 502 | __func__, hdr->inode->i_ino, |
dc70d7b3 AA |
503 | data->args.pgbase, (size_t)data->args.count, offset); |
504 | ||
505 | /* Retrieve the correct rpc_client for the byte range */ | |
506 | j = nfs4_fl_calc_j_index(lseg, offset); | |
507 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
508 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 509 | if (!ds) |
dc70d7b3 | 510 | return PNFS_NOT_ATTEMPTED; |
3a7936c3 AA |
511 | dprintk("%s USE DS: %s cl_count %d\n", __func__, |
512 | ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count)); | |
dc70d7b3 AA |
513 | |
514 | /* No multipath support. Use first DS */ | |
3a7936c3 | 515 | atomic_inc(&ds->ds_clp->cl_count); |
dc70d7b3 AA |
516 | data->ds_clp = ds->ds_clp; |
517 | fh = nfs4_fl_select_ds_fh(lseg, j); | |
518 | if (fh) | |
519 | data->args.fh = fh; | |
520 | ||
521 | data->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
522 | data->mds_offset = offset; | |
523 | ||
524 | /* Perform an asynchronous read to ds */ | |
c5996c4e | 525 | status = nfs_initiate_read(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 | 526 | &filelayout_read_call_ops, RPC_TASK_SOFTCONN); |
dc70d7b3 AA |
527 | BUG_ON(status != 0); |
528 | return PNFS_ATTEMPTED; | |
529 | } | |
530 | ||
a69aef14 | 531 | /* Perform async writes. */ |
0382b744 AA |
532 | static enum pnfs_try_status |
533 | filelayout_write_pagelist(struct nfs_write_data *data, int sync) | |
534 | { | |
cd841605 FI |
535 | struct nfs_pgio_header *hdr = data->header; |
536 | struct pnfs_layout_segment *lseg = hdr->lseg; | |
a69aef14 FI |
537 | struct nfs4_pnfs_ds *ds; |
538 | loff_t offset = data->args.offset; | |
539 | u32 j, idx; | |
540 | struct nfs_fh *fh; | |
541 | int status; | |
542 | ||
543 | /* Retrieve the correct rpc_client for the byte range */ | |
544 | j = nfs4_fl_calc_j_index(lseg, offset); | |
545 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
546 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 547 | if (!ds) |
a69aef14 | 548 | return PNFS_NOT_ATTEMPTED; |
3a7936c3 AA |
549 | dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n", |
550 | __func__, hdr->inode->i_ino, sync, (size_t) data->args.count, | |
551 | offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count)); | |
a69aef14 | 552 | |
a69aef14 | 553 | data->write_done_cb = filelayout_write_done_cb; |
3a7936c3 | 554 | atomic_inc(&ds->ds_clp->cl_count); |
a69aef14 FI |
555 | data->ds_clp = ds->ds_clp; |
556 | fh = nfs4_fl_select_ds_fh(lseg, j); | |
557 | if (fh) | |
558 | data->args.fh = fh; | |
559 | /* | |
560 | * Get the file offset on the dserver. Set the write offset to | |
561 | * this offset and save the original offset. | |
562 | */ | |
563 | data->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
a69aef14 FI |
564 | |
565 | /* Perform an asynchronous write */ | |
c5996c4e | 566 | status = nfs_initiate_write(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 AA |
567 | &filelayout_write_call_ops, sync, |
568 | RPC_TASK_SOFTCONN); | |
a69aef14 FI |
569 | BUG_ON(status != 0); |
570 | return PNFS_ATTEMPTED; | |
0382b744 AA |
571 | } |
572 | ||
16b374ca AA |
573 | /* |
574 | * filelayout_check_layout() | |
575 | * | |
576 | * Make sure layout segment parameters are sane WRT the device. | |
577 | * At this point no generic layer initialization of the lseg has occurred, | |
578 | * and nothing has been added to the layout_hdr cache. | |
579 | * | |
580 | */ | |
581 | static int | |
582 | filelayout_check_layout(struct pnfs_layout_hdr *lo, | |
583 | struct nfs4_filelayout_segment *fl, | |
584 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 TM |
585 | struct nfs4_deviceid *id, |
586 | gfp_t gfp_flags) | |
16b374ca | 587 | { |
a1eaecbc | 588 | struct nfs4_deviceid_node *d; |
16b374ca AA |
589 | struct nfs4_file_layout_dsaddr *dsaddr; |
590 | int status = -EINVAL; | |
b7edfaa1 | 591 | struct nfs_server *nfss = NFS_SERVER(lo->plh_inode); |
16b374ca AA |
592 | |
593 | dprintk("--> %s\n", __func__); | |
594 | ||
7c24d948 AA |
595 | /* FIXME: remove this check when layout segment support is added */ |
596 | if (lgr->range.offset != 0 || | |
597 | lgr->range.length != NFS4_MAX_UINT64) { | |
598 | dprintk("%s Only whole file layouts supported. Use MDS i/o\n", | |
599 | __func__); | |
600 | goto out; | |
601 | } | |
602 | ||
16b374ca | 603 | if (fl->pattern_offset > lgr->range.offset) { |
3b6445a6 | 604 | dprintk("%s pattern_offset %lld too large\n", |
16b374ca AA |
605 | __func__, fl->pattern_offset); |
606 | goto out; | |
607 | } | |
608 | ||
75247aff BH |
609 | if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) { |
610 | dprintk("%s Invalid stripe unit (%u)\n", | |
16b374ca AA |
611 | __func__, fl->stripe_unit); |
612 | goto out; | |
613 | } | |
614 | ||
615 | /* find and reference the deviceid */ | |
35c8bb54 BH |
616 | d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld, |
617 | NFS_SERVER(lo->plh_inode)->nfs_client, id); | |
a1eaecbc | 618 | if (d == NULL) { |
a75b9df9 | 619 | dsaddr = get_device_info(lo->plh_inode, id, gfp_flags); |
16b374ca AA |
620 | if (dsaddr == NULL) |
621 | goto out; | |
a1eaecbc BH |
622 | } else |
623 | dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node); | |
c47abcf8 AA |
624 | /* Found deviceid is being reaped */ |
625 | if (test_bit(NFS_DEVICEID_INVALID, &dsaddr->id_node.flags)) | |
626 | goto out_put; | |
627 | ||
16b374ca AA |
628 | fl->dsaddr = dsaddr; |
629 | ||
e414966b CL |
630 | if (fl->first_stripe_index >= dsaddr->stripe_count) { |
631 | dprintk("%s Bad first_stripe_index %u\n", | |
16b374ca AA |
632 | __func__, fl->first_stripe_index); |
633 | goto out_put; | |
634 | } | |
635 | ||
636 | if ((fl->stripe_type == STRIPE_SPARSE && | |
637 | fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) || | |
638 | (fl->stripe_type == STRIPE_DENSE && | |
639 | fl->num_fh != dsaddr->stripe_count)) { | |
640 | dprintk("%s num_fh %u not valid for given packing\n", | |
641 | __func__, fl->num_fh); | |
642 | goto out_put; | |
643 | } | |
644 | ||
645 | if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) { | |
646 | dprintk("%s Stripe unit (%u) not aligned with rsize %u " | |
647 | "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize, | |
648 | nfss->wsize); | |
649 | } | |
650 | ||
651 | status = 0; | |
652 | out: | |
653 | dprintk("--> %s returns %d\n", __func__, status); | |
654 | return status; | |
655 | out_put: | |
ea8eecdd | 656 | nfs4_fl_put_deviceid(dsaddr); |
16b374ca AA |
657 | goto out; |
658 | } | |
659 | ||
660 | static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl) | |
661 | { | |
662 | int i; | |
663 | ||
664 | for (i = 0; i < fl->num_fh; i++) { | |
665 | if (!fl->fh_array[i]) | |
666 | break; | |
667 | kfree(fl->fh_array[i]); | |
668 | } | |
669 | kfree(fl->fh_array); | |
670 | fl->fh_array = NULL; | |
671 | } | |
672 | ||
673 | static void | |
674 | _filelayout_free_lseg(struct nfs4_filelayout_segment *fl) | |
675 | { | |
676 | filelayout_free_fh_array(fl); | |
677 | kfree(fl); | |
678 | } | |
679 | ||
680 | static int | |
681 | filelayout_decode_layout(struct pnfs_layout_hdr *flo, | |
682 | struct nfs4_filelayout_segment *fl, | |
683 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 TM |
684 | struct nfs4_deviceid *id, |
685 | gfp_t gfp_flags) | |
16b374ca | 686 | { |
35124a09 | 687 | struct xdr_stream stream; |
f7da7a12 | 688 | struct xdr_buf buf; |
35124a09 WAA |
689 | struct page *scratch; |
690 | __be32 *p; | |
16b374ca AA |
691 | uint32_t nfl_util; |
692 | int i; | |
693 | ||
694 | dprintk("%s: set_layout_map Begin\n", __func__); | |
695 | ||
a75b9df9 | 696 | scratch = alloc_page(gfp_flags); |
35124a09 WAA |
697 | if (!scratch) |
698 | return -ENOMEM; | |
699 | ||
f7da7a12 | 700 | xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len); |
35124a09 WAA |
701 | xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE); |
702 | ||
703 | /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8), | |
704 | * num_fh (4) */ | |
705 | p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20); | |
706 | if (unlikely(!p)) | |
707 | goto out_err; | |
708 | ||
16b374ca AA |
709 | memcpy(id, p, sizeof(*id)); |
710 | p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE); | |
a1eaecbc | 711 | nfs4_print_deviceid(id); |
16b374ca AA |
712 | |
713 | nfl_util = be32_to_cpup(p++); | |
714 | if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS) | |
715 | fl->commit_through_mds = 1; | |
716 | if (nfl_util & NFL4_UFLG_DENSE) | |
717 | fl->stripe_type = STRIPE_DENSE; | |
718 | else | |
719 | fl->stripe_type = STRIPE_SPARSE; | |
720 | fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK; | |
721 | ||
722 | fl->first_stripe_index = be32_to_cpup(p++); | |
723 | p = xdr_decode_hyper(p, &fl->pattern_offset); | |
724 | fl->num_fh = be32_to_cpup(p++); | |
725 | ||
726 | dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n", | |
727 | __func__, nfl_util, fl->num_fh, fl->first_stripe_index, | |
728 | fl->pattern_offset); | |
729 | ||
cec765cf AA |
730 | /* Note that a zero value for num_fh is legal for STRIPE_SPARSE. |
731 | * Futher checking is done in filelayout_check_layout */ | |
e414966b | 732 | if (fl->num_fh > |
cec765cf | 733 | max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT)) |
35124a09 WAA |
734 | goto out_err; |
735 | ||
cec765cf AA |
736 | if (fl->num_fh > 0) { |
737 | fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *), | |
738 | gfp_flags); | |
739 | if (!fl->fh_array) | |
740 | goto out_err; | |
741 | } | |
16b374ca AA |
742 | |
743 | for (i = 0; i < fl->num_fh; i++) { | |
744 | /* Do we want to use a mempool here? */ | |
a75b9df9 | 745 | fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags); |
35124a09 WAA |
746 | if (!fl->fh_array[i]) |
747 | goto out_err_free; | |
748 | ||
749 | p = xdr_inline_decode(&stream, 4); | |
750 | if (unlikely(!p)) | |
751 | goto out_err_free; | |
16b374ca AA |
752 | fl->fh_array[i]->size = be32_to_cpup(p++); |
753 | if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) { | |
f9fd2d9c | 754 | printk(KERN_ERR "NFS: Too big fh %d received %d\n", |
16b374ca | 755 | i, fl->fh_array[i]->size); |
35124a09 | 756 | goto out_err_free; |
16b374ca | 757 | } |
35124a09 WAA |
758 | |
759 | p = xdr_inline_decode(&stream, fl->fh_array[i]->size); | |
760 | if (unlikely(!p)) | |
761 | goto out_err_free; | |
16b374ca | 762 | memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size); |
16b374ca AA |
763 | dprintk("DEBUG: %s: fh len %d\n", __func__, |
764 | fl->fh_array[i]->size); | |
765 | } | |
766 | ||
35124a09 | 767 | __free_page(scratch); |
16b374ca | 768 | return 0; |
35124a09 WAA |
769 | |
770 | out_err_free: | |
771 | filelayout_free_fh_array(fl); | |
772 | out_err: | |
773 | __free_page(scratch); | |
774 | return -EIO; | |
16b374ca AA |
775 | } |
776 | ||
c879513e FI |
777 | static void |
778 | filelayout_free_lseg(struct pnfs_layout_segment *lseg) | |
779 | { | |
780 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
781 | ||
782 | dprintk("--> %s\n", __func__); | |
783 | nfs4_fl_put_deviceid(fl->dsaddr); | |
799ba8d5 FI |
784 | /* This assumes a single RW lseg */ |
785 | if (lseg->pls_range.iomode == IOMODE_RW) { | |
786 | struct nfs4_filelayout *flo; | |
787 | ||
788 | flo = FILELAYOUT_FROM_HDR(lseg->pls_layout); | |
789 | flo->commit_info.nbuckets = 0; | |
790 | kfree(flo->commit_info.buckets); | |
791 | flo->commit_info.buckets = NULL; | |
792 | } | |
c879513e FI |
793 | _filelayout_free_lseg(fl); |
794 | } | |
795 | ||
799ba8d5 FI |
796 | static int |
797 | filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg, | |
ea2cf228 | 798 | struct nfs_commit_info *cinfo, |
799ba8d5 FI |
799 | gfp_t gfp_flags) |
800 | { | |
801 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
ea2cf228 | 802 | struct pnfs_commit_bucket *buckets; |
799ba8d5 FI |
803 | int size; |
804 | ||
805 | if (fl->commit_through_mds) | |
806 | return 0; | |
ea2cf228 | 807 | if (cinfo->ds->nbuckets != 0) { |
799ba8d5 FI |
808 | /* This assumes there is only one IOMODE_RW lseg. What |
809 | * we really want to do is have a layout_hdr level | |
810 | * dictionary of <multipath_list4, fh> keys, each | |
811 | * associated with a struct list_head, populated by calls | |
812 | * to filelayout_write_pagelist(). | |
813 | * */ | |
814 | return 0; | |
815 | } | |
816 | ||
817 | size = (fl->stripe_type == STRIPE_SPARSE) ? | |
818 | fl->dsaddr->ds_num : fl->dsaddr->stripe_count; | |
819 | ||
ea2cf228 | 820 | buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket), |
799ba8d5 FI |
821 | gfp_flags); |
822 | if (!buckets) | |
823 | return -ENOMEM; | |
824 | else { | |
825 | int i; | |
826 | ||
ea2cf228 FI |
827 | spin_lock(cinfo->lock); |
828 | if (cinfo->ds->nbuckets != 0) | |
799ba8d5 FI |
829 | kfree(buckets); |
830 | else { | |
ea2cf228 FI |
831 | cinfo->ds->buckets = buckets; |
832 | cinfo->ds->nbuckets = size; | |
799ba8d5 FI |
833 | for (i = 0; i < size; i++) { |
834 | INIT_LIST_HEAD(&buckets[i].written); | |
835 | INIT_LIST_HEAD(&buckets[i].committing); | |
836 | } | |
837 | } | |
ea2cf228 | 838 | spin_unlock(cinfo->lock); |
799ba8d5 FI |
839 | return 0; |
840 | } | |
841 | } | |
842 | ||
16b374ca AA |
843 | static struct pnfs_layout_segment * |
844 | filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid, | |
a75b9df9 TM |
845 | struct nfs4_layoutget_res *lgr, |
846 | gfp_t gfp_flags) | |
16b374ca AA |
847 | { |
848 | struct nfs4_filelayout_segment *fl; | |
849 | int rc; | |
850 | struct nfs4_deviceid id; | |
851 | ||
852 | dprintk("--> %s\n", __func__); | |
a75b9df9 | 853 | fl = kzalloc(sizeof(*fl), gfp_flags); |
16b374ca AA |
854 | if (!fl) |
855 | return NULL; | |
856 | ||
a75b9df9 TM |
857 | rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags); |
858 | if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) { | |
16b374ca AA |
859 | _filelayout_free_lseg(fl); |
860 | return NULL; | |
861 | } | |
862 | return &fl->generic_hdr; | |
863 | } | |
864 | ||
94ad1c80 FI |
865 | /* |
866 | * filelayout_pg_test(). Called by nfs_can_coalesce_requests() | |
867 | * | |
18ad0a9f BH |
868 | * return true : coalesce page |
869 | * return false : don't coalesce page | |
94ad1c80 | 870 | */ |
1751c363 | 871 | static bool |
94ad1c80 FI |
872 | filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev, |
873 | struct nfs_page *req) | |
874 | { | |
875 | u64 p_stripe, r_stripe; | |
876 | u32 stripe_unit; | |
877 | ||
19345cb2 BH |
878 | if (!pnfs_generic_pg_test(pgio, prev, req) || |
879 | !nfs_generic_pg_test(pgio, prev, req)) | |
880 | return false; | |
89a58e32 | 881 | |
b5542849 FI |
882 | p_stripe = (u64)req_offset(prev); |
883 | r_stripe = (u64)req_offset(req); | |
94ad1c80 FI |
884 | stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit; |
885 | ||
886 | do_div(p_stripe, stripe_unit); | |
887 | do_div(r_stripe, stripe_unit); | |
888 | ||
889 | return (p_stripe == r_stripe); | |
890 | } | |
891 | ||
17280175 | 892 | static void |
7c24d948 AA |
893 | filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio, |
894 | struct nfs_page *req) | |
895 | { | |
896 | BUG_ON(pgio->pg_lseg != NULL); | |
897 | ||
1825a0d0 FI |
898 | if (req->wb_offset != req->wb_pgbase) { |
899 | /* | |
900 | * Handling unaligned pages is difficult, because have to | |
901 | * somehow split a req in two in certain cases in the | |
902 | * pg.test code. Avoid this by just not using pnfs | |
903 | * in this case. | |
904 | */ | |
905 | nfs_pageio_reset_read_mds(pgio); | |
906 | return; | |
907 | } | |
7c24d948 AA |
908 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
909 | req->wb_context, | |
910 | 0, | |
911 | NFS4_MAX_UINT64, | |
912 | IOMODE_READ, | |
913 | GFP_KERNEL); | |
914 | /* If no lseg, fall back to read through mds */ | |
915 | if (pgio->pg_lseg == NULL) | |
1f945357 | 916 | nfs_pageio_reset_read_mds(pgio); |
7c24d948 AA |
917 | } |
918 | ||
17280175 | 919 | static void |
7c24d948 AA |
920 | filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio, |
921 | struct nfs_page *req) | |
922 | { | |
ea2cf228 | 923 | struct nfs_commit_info cinfo; |
799ba8d5 FI |
924 | int status; |
925 | ||
7c24d948 AA |
926 | BUG_ON(pgio->pg_lseg != NULL); |
927 | ||
1825a0d0 FI |
928 | if (req->wb_offset != req->wb_pgbase) |
929 | goto out_mds; | |
7c24d948 AA |
930 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
931 | req->wb_context, | |
932 | 0, | |
933 | NFS4_MAX_UINT64, | |
934 | IOMODE_RW, | |
935 | GFP_NOFS); | |
936 | /* If no lseg, fall back to write through mds */ | |
937 | if (pgio->pg_lseg == NULL) | |
799ba8d5 | 938 | goto out_mds; |
ea2cf228 FI |
939 | nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq); |
940 | status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS); | |
799ba8d5 FI |
941 | if (status < 0) { |
942 | put_lseg(pgio->pg_lseg); | |
943 | pgio->pg_lseg = NULL; | |
944 | goto out_mds; | |
945 | } | |
946 | return; | |
947 | out_mds: | |
948 | nfs_pageio_reset_write_mds(pgio); | |
7c24d948 AA |
949 | } |
950 | ||
1751c363 | 951 | static const struct nfs_pageio_ops filelayout_pg_read_ops = { |
7c24d948 | 952 | .pg_init = filelayout_pg_init_read, |
1751c363 | 953 | .pg_test = filelayout_pg_test, |
493292dd | 954 | .pg_doio = pnfs_generic_pg_readpages, |
1751c363 TM |
955 | }; |
956 | ||
957 | static const struct nfs_pageio_ops filelayout_pg_write_ops = { | |
7c24d948 | 958 | .pg_init = filelayout_pg_init_write, |
1751c363 | 959 | .pg_test = filelayout_pg_test, |
dce81290 | 960 | .pg_doio = pnfs_generic_pg_writepages, |
1751c363 TM |
961 | }; |
962 | ||
e0c2b380 FI |
963 | static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j) |
964 | { | |
965 | if (fl->stripe_type == STRIPE_SPARSE) | |
966 | return nfs4_fl_calc_ds_index(&fl->generic_hdr, j); | |
967 | else | |
968 | return j; | |
969 | } | |
970 | ||
d6d6dc7c FI |
971 | /* The generic layer is about to remove the req from the commit list. |
972 | * If this will make the bucket empty, it will need to put the lseg reference. | |
d6d6dc7c | 973 | */ |
8dd37758 | 974 | static void |
ea2cf228 FI |
975 | filelayout_clear_request_commit(struct nfs_page *req, |
976 | struct nfs_commit_info *cinfo) | |
d6d6dc7c | 977 | { |
8dd37758 | 978 | struct pnfs_layout_segment *freeme = NULL; |
8dd37758 | 979 | |
ea2cf228 | 980 | spin_lock(cinfo->lock); |
8dd37758 TM |
981 | if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags)) |
982 | goto out; | |
ea2cf228 | 983 | cinfo->ds->nwritten--; |
d6d6dc7c | 984 | if (list_is_singular(&req->wb_list)) { |
ea2cf228 | 985 | struct pnfs_commit_bucket *bucket; |
d6d6dc7c | 986 | |
799ba8d5 | 987 | bucket = list_first_entry(&req->wb_list, |
ea2cf228 | 988 | struct pnfs_commit_bucket, |
799ba8d5 FI |
989 | written); |
990 | freeme = bucket->wlseg; | |
991 | bucket->wlseg = NULL; | |
d6d6dc7c | 992 | } |
8dd37758 | 993 | out: |
ea2cf228 FI |
994 | nfs_request_remove_commit_list(req, cinfo); |
995 | spin_unlock(cinfo->lock); | |
8dd37758 | 996 | put_lseg(freeme); |
d6d6dc7c FI |
997 | } |
998 | ||
999 | static struct list_head * | |
1000 | filelayout_choose_commit_list(struct nfs_page *req, | |
ea2cf228 FI |
1001 | struct pnfs_layout_segment *lseg, |
1002 | struct nfs_commit_info *cinfo) | |
e0c2b380 | 1003 | { |
e0c2b380 FI |
1004 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); |
1005 | u32 i, j; | |
1006 | struct list_head *list; | |
ea2cf228 | 1007 | struct pnfs_commit_bucket *buckets; |
e0c2b380 | 1008 | |
d6d6dc7c | 1009 | if (fl->commit_through_mds) |
ea2cf228 | 1010 | return &cinfo->mds->list; |
d6d6dc7c | 1011 | |
e0c2b380 FI |
1012 | /* Note that we are calling nfs4_fl_calc_j_index on each page |
1013 | * that ends up being committed to a data server. An attractive | |
1014 | * alternative is to add a field to nfs_write_data and nfs_page | |
1015 | * to store the value calculated in filelayout_write_pagelist | |
1016 | * and just use that here. | |
1017 | */ | |
b5542849 | 1018 | j = nfs4_fl_calc_j_index(lseg, req_offset(req)); |
e0c2b380 | 1019 | i = select_bucket_index(fl, j); |
ea2cf228 | 1020 | buckets = cinfo->ds->buckets; |
799ba8d5 | 1021 | list = &buckets[i].written; |
e0c2b380 | 1022 | if (list_empty(list)) { |
d6d6dc7c FI |
1023 | /* Non-empty buckets hold a reference on the lseg. That ref |
1024 | * is normally transferred to the COMMIT call and released | |
1025 | * there. It could also be released if the last req is pulled | |
1026 | * off due to a rewrite, in which case it will be done in | |
799ba8d5 | 1027 | * filelayout_clear_request_commit |
d6d6dc7c | 1028 | */ |
799ba8d5 | 1029 | buckets[i].wlseg = get_lseg(lseg); |
e0c2b380 | 1030 | } |
8dd37758 | 1031 | set_bit(PG_COMMIT_TO_DS, &req->wb_flags); |
ea2cf228 | 1032 | cinfo->ds->nwritten++; |
e0c2b380 FI |
1033 | return list; |
1034 | } | |
1035 | ||
8dd37758 TM |
1036 | static void |
1037 | filelayout_mark_request_commit(struct nfs_page *req, | |
ea2cf228 FI |
1038 | struct pnfs_layout_segment *lseg, |
1039 | struct nfs_commit_info *cinfo) | |
8dd37758 TM |
1040 | { |
1041 | struct list_head *list; | |
1042 | ||
ea2cf228 FI |
1043 | list = filelayout_choose_commit_list(req, lseg, cinfo); |
1044 | nfs_request_add_commit_list(req, list, cinfo); | |
8dd37758 TM |
1045 | } |
1046 | ||
e0c2b380 FI |
1047 | static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i) |
1048 | { | |
1049 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1050 | ||
1051 | if (flseg->stripe_type == STRIPE_SPARSE) | |
1052 | return i; | |
1053 | else | |
1054 | return nfs4_fl_calc_ds_index(lseg, i); | |
1055 | } | |
1056 | ||
1057 | static struct nfs_fh * | |
1058 | select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i) | |
1059 | { | |
1060 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1061 | ||
1062 | if (flseg->stripe_type == STRIPE_SPARSE) { | |
1063 | if (flseg->num_fh == 1) | |
1064 | i = 0; | |
1065 | else if (flseg->num_fh == 0) | |
1066 | /* Use the MDS OPEN fh set in nfs_read_rpcsetup */ | |
1067 | return NULL; | |
1068 | } | |
1069 | return flseg->fh_array[i]; | |
1070 | } | |
1071 | ||
0b7c0153 | 1072 | static int filelayout_initiate_commit(struct nfs_commit_data *data, int how) |
e0c2b380 FI |
1073 | { |
1074 | struct pnfs_layout_segment *lseg = data->lseg; | |
1075 | struct nfs4_pnfs_ds *ds; | |
1076 | u32 idx; | |
1077 | struct nfs_fh *fh; | |
1078 | ||
1079 | idx = calc_ds_index_from_commit(lseg, data->ds_commit_index); | |
1080 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
1081 | if (!ds) { | |
e0c2b380 | 1082 | prepare_to_resend_writes(data); |
8dd37758 | 1083 | filelayout_commit_release(data); |
e0c2b380 FI |
1084 | return -EAGAIN; |
1085 | } | |
3a7936c3 AA |
1086 | dprintk("%s ino %lu, how %d cl_count %d\n", __func__, |
1087 | data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count)); | |
0b7c0153 | 1088 | data->commit_done_cb = filelayout_commit_done_cb; |
3a7936c3 | 1089 | atomic_inc(&ds->ds_clp->cl_count); |
e0c2b380 FI |
1090 | data->ds_clp = ds->ds_clp; |
1091 | fh = select_ds_fh_from_commit(lseg, data->ds_commit_index); | |
1092 | if (fh) | |
1093 | data->args.fh = fh; | |
0b7c0153 | 1094 | return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 AA |
1095 | &filelayout_commit_call_ops, how, |
1096 | RPC_TASK_SOFTCONN); | |
e0c2b380 FI |
1097 | } |
1098 | ||
8dd37758 | 1099 | static int |
1763da12 FI |
1100 | transfer_commit_list(struct list_head *src, struct list_head *dst, |
1101 | struct nfs_commit_info *cinfo, int max) | |
8dd37758 | 1102 | { |
8dd37758 TM |
1103 | struct nfs_page *req, *tmp; |
1104 | int ret = 0; | |
1105 | ||
1106 | list_for_each_entry_safe(req, tmp, src, wb_list) { | |
1107 | if (!nfs_lock_request(req)) | |
1108 | continue; | |
ea2cf228 | 1109 | if (cond_resched_lock(cinfo->lock)) |
3b3be88d | 1110 | list_safe_reset_next(req, tmp, wb_list); |
ea2cf228 | 1111 | nfs_request_remove_commit_list(req, cinfo); |
8dd37758 TM |
1112 | clear_bit(PG_COMMIT_TO_DS, &req->wb_flags); |
1113 | nfs_list_add_request(req, dst); | |
1114 | ret++; | |
1763da12 | 1115 | if ((ret == max) && !cinfo->dreq) |
8dd37758 TM |
1116 | break; |
1117 | } | |
1763da12 FI |
1118 | return ret; |
1119 | } | |
1120 | ||
1121 | static int | |
1122 | filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket, | |
1123 | struct nfs_commit_info *cinfo, | |
1124 | int max) | |
1125 | { | |
1126 | struct list_head *src = &bucket->written; | |
1127 | struct list_head *dst = &bucket->committing; | |
1128 | int ret; | |
1129 | ||
1130 | ret = transfer_commit_list(src, dst, cinfo, max); | |
799ba8d5 | 1131 | if (ret) { |
ea2cf228 FI |
1132 | cinfo->ds->nwritten -= ret; |
1133 | cinfo->ds->ncommitting += ret; | |
799ba8d5 FI |
1134 | bucket->clseg = bucket->wlseg; |
1135 | if (list_empty(src)) | |
1136 | bucket->wlseg = NULL; | |
1137 | else | |
1138 | get_lseg(bucket->clseg); | |
1139 | } | |
8dd37758 TM |
1140 | return ret; |
1141 | } | |
1142 | ||
d6d6dc7c | 1143 | /* Move reqs from written to committing lists, returning count of number moved. |
ea2cf228 | 1144 | * Note called with cinfo->lock held. |
d6d6dc7c | 1145 | */ |
ea2cf228 FI |
1146 | static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo, |
1147 | int max) | |
d6d6dc7c | 1148 | { |
d6d6dc7c FI |
1149 | int i, rv = 0, cnt; |
1150 | ||
ea2cf228 FI |
1151 | for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) { |
1152 | cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i], | |
1153 | cinfo, max); | |
d6d6dc7c FI |
1154 | max -= cnt; |
1155 | rv += cnt; | |
1156 | } | |
d6d6dc7c FI |
1157 | return rv; |
1158 | } | |
1159 | ||
1763da12 FI |
1160 | /* Pull everything off the committing lists and dump into @dst */ |
1161 | static void filelayout_recover_commit_reqs(struct list_head *dst, | |
1162 | struct nfs_commit_info *cinfo) | |
1163 | { | |
1164 | struct pnfs_commit_bucket *b; | |
1165 | int i; | |
1166 | ||
1167 | /* NOTE cinfo->lock is NOT held, relying on fact that this is | |
1168 | * only called on single thread per dreq. | |
1169 | * Can't take the lock because need to do put_lseg | |
1170 | */ | |
1171 | for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) { | |
1172 | if (transfer_commit_list(&b->written, dst, cinfo, 0)) { | |
1173 | BUG_ON(!list_empty(&b->written)); | |
1174 | put_lseg(b->wlseg); | |
1175 | b->wlseg = NULL; | |
1176 | } | |
1177 | } | |
1178 | cinfo->ds->nwritten = 0; | |
1179 | } | |
1180 | ||
9390f425 | 1181 | static unsigned int |
ea2cf228 | 1182 | alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list) |
e0c2b380 | 1183 | { |
ea2cf228 FI |
1184 | struct pnfs_ds_commit_info *fl_cinfo; |
1185 | struct pnfs_commit_bucket *bucket; | |
0b7c0153 | 1186 | struct nfs_commit_data *data; |
e0c2b380 | 1187 | int i, j; |
9390f425 | 1188 | unsigned int nreq = 0; |
e0c2b380 | 1189 | |
ea2cf228 | 1190 | fl_cinfo = cinfo->ds; |
799ba8d5 FI |
1191 | bucket = fl_cinfo->buckets; |
1192 | for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) { | |
1193 | if (list_empty(&bucket->committing)) | |
e0c2b380 FI |
1194 | continue; |
1195 | data = nfs_commitdata_alloc(); | |
1196 | if (!data) | |
9390f425 | 1197 | break; |
e0c2b380 | 1198 | data->ds_commit_index = i; |
799ba8d5 FI |
1199 | data->lseg = bucket->clseg; |
1200 | bucket->clseg = NULL; | |
e0c2b380 | 1201 | list_add(&data->pages, list); |
9390f425 | 1202 | nreq++; |
e0c2b380 | 1203 | } |
e0c2b380 | 1204 | |
9390f425 | 1205 | /* Clean up on error */ |
799ba8d5 FI |
1206 | for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) { |
1207 | if (list_empty(&bucket->committing)) | |
e0c2b380 | 1208 | continue; |
ea2cf228 | 1209 | nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo); |
799ba8d5 FI |
1210 | put_lseg(bucket->clseg); |
1211 | bucket->clseg = NULL; | |
e0c2b380 | 1212 | } |
e0c2b380 | 1213 | /* Caller will clean up entries put on list */ |
9390f425 | 1214 | return nreq; |
e0c2b380 FI |
1215 | } |
1216 | ||
1217 | /* This follows nfs_commit_list pretty closely */ | |
1218 | static int | |
1219 | filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages, | |
ea2cf228 | 1220 | int how, struct nfs_commit_info *cinfo) |
e0c2b380 | 1221 | { |
0b7c0153 | 1222 | struct nfs_commit_data *data, *tmp; |
e0c2b380 | 1223 | LIST_HEAD(list); |
9390f425 | 1224 | unsigned int nreq = 0; |
e0c2b380 FI |
1225 | |
1226 | if (!list_empty(mds_pages)) { | |
1227 | data = nfs_commitdata_alloc(); | |
9390f425 TM |
1228 | if (data != NULL) { |
1229 | data->lseg = NULL; | |
1230 | list_add(&data->pages, &list); | |
1231 | nreq++; | |
1232 | } else | |
ea2cf228 | 1233 | nfs_retry_commit(mds_pages, NULL, cinfo); |
e0c2b380 FI |
1234 | } |
1235 | ||
ea2cf228 | 1236 | nreq += alloc_ds_commits(cinfo, &list); |
9390f425 TM |
1237 | |
1238 | if (nreq == 0) { | |
f453a54a | 1239 | cinfo->completion_ops->error_cleanup(NFS_I(inode)); |
9390f425 TM |
1240 | goto out; |
1241 | } | |
1242 | ||
ea2cf228 | 1243 | atomic_add(nreq, &cinfo->mds->rpcs_out); |
e0c2b380 FI |
1244 | |
1245 | list_for_each_entry_safe(data, tmp, &list, pages) { | |
1246 | list_del_init(&data->pages); | |
e0c2b380 | 1247 | if (!data->lseg) { |
f453a54a | 1248 | nfs_init_commit(data, mds_pages, NULL, cinfo); |
0b7c0153 | 1249 | nfs_initiate_commit(NFS_CLIENT(inode), data, |
9f0ec176 | 1250 | data->mds_ops, how, 0); |
e0c2b380 | 1251 | } else { |
ea2cf228 | 1252 | struct pnfs_commit_bucket *buckets; |
799ba8d5 | 1253 | |
ea2cf228 | 1254 | buckets = cinfo->ds->buckets; |
f453a54a | 1255 | nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo); |
e0c2b380 FI |
1256 | filelayout_initiate_commit(data, how); |
1257 | } | |
1258 | } | |
9390f425 | 1259 | out: |
ea2cf228 | 1260 | cinfo->ds->ncommitting = 0; |
9390f425 | 1261 | return PNFS_ATTEMPTED; |
e0c2b380 FI |
1262 | } |
1263 | ||
1775bc34 BH |
1264 | static void |
1265 | filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d) | |
1266 | { | |
1267 | nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node)); | |
1268 | } | |
1269 | ||
799ba8d5 FI |
1270 | static struct pnfs_layout_hdr * |
1271 | filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags) | |
1272 | { | |
1273 | struct nfs4_filelayout *flo; | |
1274 | ||
1275 | flo = kzalloc(sizeof(*flo), gfp_flags); | |
1276 | return &flo->generic_hdr; | |
1277 | } | |
1278 | ||
1279 | static void | |
1280 | filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo) | |
1281 | { | |
1282 | kfree(FILELAYOUT_FROM_HDR(lo)); | |
1283 | } | |
1284 | ||
ea2cf228 FI |
1285 | static struct pnfs_ds_commit_info * |
1286 | filelayout_get_ds_info(struct inode *inode) | |
1287 | { | |
df011748 FI |
1288 | struct pnfs_layout_hdr *layout = NFS_I(inode)->layout; |
1289 | ||
1290 | if (layout == NULL) | |
1291 | return NULL; | |
1292 | else | |
1293 | return &FILELAYOUT_FROM_HDR(layout)->commit_info; | |
ea2cf228 FI |
1294 | } |
1295 | ||
7ab672ce | 1296 | static struct pnfs_layoutdriver_type filelayout_type = { |
ea8eecdd CH |
1297 | .id = LAYOUT_NFSV4_1_FILES, |
1298 | .name = "LAYOUT_NFSV4_1_FILES", | |
1299 | .owner = THIS_MODULE, | |
799ba8d5 FI |
1300 | .alloc_layout_hdr = filelayout_alloc_layout_hdr, |
1301 | .free_layout_hdr = filelayout_free_layout_hdr, | |
ea8eecdd CH |
1302 | .alloc_lseg = filelayout_alloc_lseg, |
1303 | .free_lseg = filelayout_free_lseg, | |
1751c363 TM |
1304 | .pg_read_ops = &filelayout_pg_read_ops, |
1305 | .pg_write_ops = &filelayout_pg_write_ops, | |
ea2cf228 | 1306 | .get_ds_info = &filelayout_get_ds_info, |
8dd37758 TM |
1307 | .mark_request_commit = filelayout_mark_request_commit, |
1308 | .clear_request_commit = filelayout_clear_request_commit, | |
d6d6dc7c | 1309 | .scan_commit_lists = filelayout_scan_commit_lists, |
1763da12 | 1310 | .recover_commit_reqs = filelayout_recover_commit_reqs, |
e0c2b380 | 1311 | .commit_pagelist = filelayout_commit_pagelist, |
dc70d7b3 | 1312 | .read_pagelist = filelayout_read_pagelist, |
0382b744 | 1313 | .write_pagelist = filelayout_write_pagelist, |
1775bc34 | 1314 | .free_deviceid_node = filelayout_free_deveiceid_node, |
7ab672ce DH |
1315 | }; |
1316 | ||
1317 | static int __init nfs4filelayout_init(void) | |
1318 | { | |
1319 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n", | |
1320 | __func__); | |
1321 | return pnfs_register_layoutdriver(&filelayout_type); | |
1322 | } | |
1323 | ||
1324 | static void __exit nfs4filelayout_exit(void) | |
1325 | { | |
1326 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n", | |
1327 | __func__); | |
1328 | pnfs_unregister_layoutdriver(&filelayout_type); | |
1329 | } | |
1330 | ||
f85ef69c BF |
1331 | MODULE_ALIAS("nfs-layouttype4-1"); |
1332 | ||
7ab672ce DH |
1333 | module_init(nfs4filelayout_init); |
1334 | module_exit(nfs4filelayout_exit); |