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; | |
e7dd79af AA |
88 | struct rpc_task *task = &data->task; |
89 | ||
90 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { | |
91 | dprintk("%s Reset task %5u for i/o through MDS " | |
92 | "(req %s/%lld, %u bytes @ offset %llu)\n", __func__, | |
93 | data->task.tk_pid, | |
497826af BS |
94 | hdr->inode->i_sb->s_id, |
95 | (long long)NFS_FILEID(hdr->inode), | |
e7dd79af AA |
96 | data->args.count, |
97 | (unsigned long long)data->args.offset); | |
98 | ||
99 | task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode, | |
100 | &hdr->pages, | |
101 | hdr->completion_ops); | |
102 | } | |
103 | } | |
104 | ||
105 | static void filelayout_reset_read(struct nfs_read_data *data) | |
106 | { | |
107 | struct nfs_pgio_header *hdr = data->header; | |
e7dd79af AA |
108 | struct rpc_task *task = &data->task; |
109 | ||
110 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { | |
111 | dprintk("%s Reset task %5u for i/o through MDS " | |
112 | "(req %s/%lld, %u bytes @ offset %llu)\n", __func__, | |
113 | data->task.tk_pid, | |
497826af BS |
114 | hdr->inode->i_sb->s_id, |
115 | (long long)NFS_FILEID(hdr->inode), | |
e7dd79af AA |
116 | data->args.count, |
117 | (unsigned long long)data->args.offset); | |
118 | ||
119 | task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode, | |
120 | &hdr->pages, | |
121 | hdr->completion_ops); | |
122 | } | |
123 | } | |
124 | ||
d527e5c1 TM |
125 | static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo) |
126 | { | |
127 | if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) | |
128 | return; | |
129 | clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags); | |
130 | pnfs_return_layout(inode); | |
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 | { |
d527e5c1 TM |
138 | struct pnfs_layout_hdr *lo = lseg->pls_layout; |
139 | struct inode *inode = lo->plh_inode; | |
e7dd79af AA |
140 | struct nfs_server *mds_server = NFS_SERVER(inode); |
141 | struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg); | |
9cb81968 | 142 | struct nfs_client *mds_client = mds_server->nfs_client; |
671fb896 | 143 | struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table; |
9cb81968 | 144 | |
cbdabc7f AA |
145 | if (task->tk_status >= 0) |
146 | return 0; | |
cbdabc7f AA |
147 | |
148 | switch (task->tk_status) { | |
9cb81968 AA |
149 | /* MDS state errors */ |
150 | case -NFS4ERR_DELEG_REVOKED: | |
151 | case -NFS4ERR_ADMIN_REVOKED: | |
152 | case -NFS4ERR_BAD_STATEID: | |
e7dd79af AA |
153 | if (state == NULL) |
154 | break; | |
2dc31756 | 155 | nfs_remove_bad_delegation(state->inode); |
9cb81968 | 156 | case -NFS4ERR_OPENMODE: |
e7dd79af AA |
157 | if (state == NULL) |
158 | break; | |
9cb81968 AA |
159 | nfs4_schedule_stateid_recovery(mds_server, state); |
160 | goto wait_on_recovery; | |
161 | case -NFS4ERR_EXPIRED: | |
e7dd79af AA |
162 | if (state != NULL) |
163 | nfs4_schedule_stateid_recovery(mds_server, state); | |
9cb81968 AA |
164 | nfs4_schedule_lease_recovery(mds_client); |
165 | goto wait_on_recovery; | |
166 | /* DS session errors */ | |
cbdabc7f AA |
167 | case -NFS4ERR_BADSESSION: |
168 | case -NFS4ERR_BADSLOT: | |
169 | case -NFS4ERR_BAD_HIGH_SLOT: | |
170 | case -NFS4ERR_DEADSESSION: | |
171 | case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION: | |
172 | case -NFS4ERR_SEQ_FALSE_RETRY: | |
173 | case -NFS4ERR_SEQ_MISORDERED: | |
174 | dprintk("%s ERROR %d, Reset session. Exchangeid " | |
175 | "flags 0x%x\n", __func__, task->tk_status, | |
176 | clp->cl_exchange_flags); | |
9f594791 | 177 | nfs4_schedule_session_recovery(clp->cl_session, task->tk_status); |
cbdabc7f AA |
178 | break; |
179 | case -NFS4ERR_DELAY: | |
180 | case -NFS4ERR_GRACE: | |
181 | case -EKEYEXPIRED: | |
182 | rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX); | |
183 | break; | |
a8a4ae3a AA |
184 | case -NFS4ERR_RETRY_UNCACHED_REP: |
185 | break; | |
041245c8 AA |
186 | /* Invalidate Layout errors */ |
187 | case -NFS4ERR_PNFS_NO_LAYOUT: | |
188 | case -ESTALE: /* mapped NFS4ERR_STALE */ | |
189 | case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */ | |
190 | case -EISDIR: /* mapped NFS4ERR_ISDIR */ | |
191 | case -NFS4ERR_FHEXPIRED: | |
192 | case -NFS4ERR_WRONG_TYPE: | |
193 | dprintk("%s Invalid layout error %d\n", __func__, | |
194 | task->tk_status); | |
195 | /* | |
196 | * Destroy layout so new i/o will get a new layout. | |
197 | * Layout will not be destroyed until all current lseg | |
198 | * references are put. Mark layout as invalid to resend failed | |
199 | * i/o and all i/o waiting on the slot table to the MDS until | |
200 | * layout is destroyed and a new valid layout is obtained. | |
201 | */ | |
d42e7873 | 202 | pnfs_destroy_layout(NFS_I(inode)); |
041245c8 AA |
203 | rpc_wake_up(&tbl->slot_tbl_waitq); |
204 | goto reset; | |
e7dd79af AA |
205 | /* RPC connection errors */ |
206 | case -ECONNREFUSED: | |
207 | case -EHOSTDOWN: | |
208 | case -EHOSTUNREACH: | |
209 | case -ENETUNREACH: | |
210 | case -EIO: | |
211 | case -ETIMEDOUT: | |
212 | case -EPIPE: | |
213 | dprintk("%s DS connection error %d\n", __func__, | |
214 | task->tk_status); | |
1dfed273 | 215 | nfs4_mark_deviceid_unavailable(devid); |
d527e5c1 | 216 | set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags); |
671fb896 | 217 | rpc_wake_up(&tbl->slot_tbl_waitq); |
e7dd79af | 218 | /* fall through */ |
cbdabc7f | 219 | default: |
041245c8 | 220 | reset: |
e7dd79af | 221 | dprintk("%s Retry through MDS. Error %d\n", __func__, |
cbdabc7f | 222 | task->tk_status); |
e7dd79af | 223 | return -NFS4ERR_RESET_TO_MDS; |
cbdabc7f | 224 | } |
9cb81968 | 225 | out: |
cbdabc7f AA |
226 | task->tk_status = 0; |
227 | return -EAGAIN; | |
9cb81968 AA |
228 | wait_on_recovery: |
229 | rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL); | |
230 | if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0) | |
231 | rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task); | |
232 | goto out; | |
cbdabc7f AA |
233 | } |
234 | ||
235 | /* NFS_PROTO call done callback routines */ | |
236 | ||
237 | static int filelayout_read_done_cb(struct rpc_task *task, | |
238 | struct nfs_read_data *data) | |
239 | { | |
e7dd79af AA |
240 | struct nfs_pgio_header *hdr = data->header; |
241 | int err; | |
cbdabc7f | 242 | |
e7dd79af AA |
243 | err = filelayout_async_handle_error(task, data->args.context->state, |
244 | data->ds_clp, hdr->lseg); | |
cbdabc7f | 245 | |
e7dd79af AA |
246 | switch (err) { |
247 | case -NFS4ERR_RESET_TO_MDS: | |
248 | filelayout_reset_read(data); | |
249 | return task->tk_status; | |
250 | case -EAGAIN: | |
d00c5d43 | 251 | rpc_restart_call_prepare(task); |
cbdabc7f AA |
252 | return -EAGAIN; |
253 | } | |
254 | ||
255 | return 0; | |
256 | } | |
257 | ||
863a3c6c AA |
258 | /* |
259 | * We reference the rpc_cred of the first WRITE that triggers the need for | |
260 | * a LAYOUTCOMMIT, and use it to send the layoutcommit compound. | |
261 | * rfc5661 is not clear about which credential should be used. | |
262 | */ | |
263 | static void | |
264 | filelayout_set_layoutcommit(struct nfs_write_data *wdata) | |
265 | { | |
cd841605 FI |
266 | struct nfs_pgio_header *hdr = wdata->header; |
267 | ||
268 | if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds || | |
863a3c6c AA |
269 | wdata->res.verf->committed == NFS_FILE_SYNC) |
270 | return; | |
271 | ||
272 | pnfs_set_layoutcommit(wdata); | |
cd841605 FI |
273 | dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino, |
274 | (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb); | |
863a3c6c AA |
275 | } |
276 | ||
1dfed273 TM |
277 | bool |
278 | filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node) | |
279 | { | |
280 | return filelayout_test_devid_invalid(node) || | |
281 | nfs4_test_deviceid_unavailable(node); | |
282 | } | |
283 | ||
284 | static bool | |
285 | filelayout_reset_to_mds(struct pnfs_layout_segment *lseg) | |
286 | { | |
287 | struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg); | |
288 | ||
8006bfba | 289 | return filelayout_test_devid_unavailable(node); |
1dfed273 TM |
290 | } |
291 | ||
dc70d7b3 AA |
292 | /* |
293 | * Call ops for the async read/write cases | |
294 | * In the case of dense layouts, the offset needs to be reset to its | |
295 | * original value. | |
296 | */ | |
297 | static void filelayout_read_prepare(struct rpc_task *task, void *data) | |
298 | { | |
cd12ae32 | 299 | struct nfs_read_data *rdata = data; |
dc70d7b3 | 300 | |
041245c8 | 301 | if (filelayout_reset_to_mds(rdata->header->lseg)) { |
0ad2f378 AA |
302 | dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid); |
303 | filelayout_reset_read(rdata); | |
304 | rpc_exit(task, 0); | |
305 | return; | |
306 | } | |
cbdabc7f AA |
307 | rdata->read_done_cb = filelayout_read_done_cb; |
308 | ||
dc70d7b3 AA |
309 | if (nfs41_setup_sequence(rdata->ds_clp->cl_session, |
310 | &rdata->args.seq_args, &rdata->res.seq_res, | |
9d12b216 | 311 | task)) |
dc70d7b3 AA |
312 | return; |
313 | ||
314 | rpc_call_start(task); | |
315 | } | |
316 | ||
317 | static void filelayout_read_call_done(struct rpc_task *task, void *data) | |
318 | { | |
cd12ae32 | 319 | struct nfs_read_data *rdata = data; |
dc70d7b3 AA |
320 | |
321 | dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status); | |
322 | ||
bd4aeffb AA |
323 | if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) && |
324 | task->tk_status == 0) | |
0ad2f378 AA |
325 | return; |
326 | ||
dc70d7b3 | 327 | /* Note this may cause RPC to be resent */ |
cd841605 | 328 | rdata->header->mds_ops->rpc_call_done(task, data); |
dc70d7b3 AA |
329 | } |
330 | ||
0a702195 WAA |
331 | static void filelayout_read_count_stats(struct rpc_task *task, void *data) |
332 | { | |
cd12ae32 | 333 | struct nfs_read_data *rdata = data; |
0a702195 | 334 | |
cd841605 | 335 | rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics); |
0a702195 WAA |
336 | } |
337 | ||
dc70d7b3 AA |
338 | static void filelayout_read_release(void *data) |
339 | { | |
cd12ae32 | 340 | struct nfs_read_data *rdata = data; |
d527e5c1 | 341 | struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout; |
dc70d7b3 | 342 | |
d527e5c1 | 343 | filelayout_fenceme(lo->plh_inode, lo); |
996074cb | 344 | nfs_put_client(rdata->ds_clp); |
cd841605 | 345 | rdata->header->mds_ops->rpc_release(data); |
dc70d7b3 AA |
346 | } |
347 | ||
a69aef14 FI |
348 | static int filelayout_write_done_cb(struct rpc_task *task, |
349 | struct nfs_write_data *data) | |
350 | { | |
e7dd79af AA |
351 | struct nfs_pgio_header *hdr = data->header; |
352 | int err; | |
353 | ||
354 | err = filelayout_async_handle_error(task, data->args.context->state, | |
355 | data->ds_clp, hdr->lseg); | |
356 | ||
357 | switch (err) { | |
358 | case -NFS4ERR_RESET_TO_MDS: | |
359 | filelayout_reset_write(data); | |
360 | return task->tk_status; | |
361 | case -EAGAIN: | |
d00c5d43 | 362 | rpc_restart_call_prepare(task); |
a69aef14 FI |
363 | return -EAGAIN; |
364 | } | |
365 | ||
863a3c6c | 366 | filelayout_set_layoutcommit(data); |
a69aef14 FI |
367 | return 0; |
368 | } | |
369 | ||
e0c2b380 | 370 | /* Fake up some data that will cause nfs_commit_release to retry the writes. */ |
0b7c0153 | 371 | static void prepare_to_resend_writes(struct nfs_commit_data *data) |
e0c2b380 FI |
372 | { |
373 | struct nfs_page *first = nfs_list_entry(data->pages.next); | |
374 | ||
375 | data->task.tk_status = 0; | |
2f2c63bc TM |
376 | memcpy(&data->verf.verifier, &first->wb_verf, |
377 | sizeof(data->verf.verifier)); | |
378 | data->verf.verifier.data[0]++; /* ensure verifier mismatch */ | |
e0c2b380 FI |
379 | } |
380 | ||
381 | static int filelayout_commit_done_cb(struct rpc_task *task, | |
0b7c0153 | 382 | struct nfs_commit_data *data) |
e0c2b380 | 383 | { |
e7dd79af AA |
384 | int err; |
385 | ||
386 | err = filelayout_async_handle_error(task, NULL, data->ds_clp, | |
387 | data->lseg); | |
388 | ||
389 | switch (err) { | |
390 | case -NFS4ERR_RESET_TO_MDS: | |
391 | prepare_to_resend_writes(data); | |
392 | return -EAGAIN; | |
393 | case -EAGAIN: | |
394 | rpc_restart_call_prepare(task); | |
e0c2b380 FI |
395 | return -EAGAIN; |
396 | } | |
397 | ||
398 | return 0; | |
399 | } | |
400 | ||
a69aef14 FI |
401 | static void filelayout_write_prepare(struct rpc_task *task, void *data) |
402 | { | |
cd12ae32 | 403 | struct nfs_write_data *wdata = data; |
a69aef14 | 404 | |
041245c8 | 405 | if (filelayout_reset_to_mds(wdata->header->lseg)) { |
0ad2f378 AA |
406 | dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid); |
407 | filelayout_reset_write(wdata); | |
408 | rpc_exit(task, 0); | |
409 | return; | |
410 | } | |
a69aef14 FI |
411 | if (nfs41_setup_sequence(wdata->ds_clp->cl_session, |
412 | &wdata->args.seq_args, &wdata->res.seq_res, | |
9d12b216 | 413 | task)) |
a69aef14 FI |
414 | return; |
415 | ||
416 | rpc_call_start(task); | |
417 | } | |
418 | ||
419 | static void filelayout_write_call_done(struct rpc_task *task, void *data) | |
420 | { | |
cd12ae32 | 421 | struct nfs_write_data *wdata = data; |
a69aef14 | 422 | |
bd4aeffb AA |
423 | if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) && |
424 | task->tk_status == 0) | |
0ad2f378 AA |
425 | return; |
426 | ||
a69aef14 | 427 | /* Note this may cause RPC to be resent */ |
cd841605 | 428 | wdata->header->mds_ops->rpc_call_done(task, data); |
a69aef14 FI |
429 | } |
430 | ||
0a702195 WAA |
431 | static void filelayout_write_count_stats(struct rpc_task *task, void *data) |
432 | { | |
cd12ae32 | 433 | struct nfs_write_data *wdata = data; |
0a702195 | 434 | |
cd841605 | 435 | rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics); |
0a702195 WAA |
436 | } |
437 | ||
a69aef14 FI |
438 | static void filelayout_write_release(void *data) |
439 | { | |
cd12ae32 | 440 | struct nfs_write_data *wdata = data; |
d527e5c1 | 441 | struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout; |
a69aef14 | 442 | |
d527e5c1 | 443 | filelayout_fenceme(lo->plh_inode, lo); |
996074cb | 444 | nfs_put_client(wdata->ds_clp); |
cd841605 | 445 | wdata->header->mds_ops->rpc_release(data); |
a69aef14 FI |
446 | } |
447 | ||
0b7c0153 | 448 | static void filelayout_commit_prepare(struct rpc_task *task, void *data) |
e0c2b380 | 449 | { |
0b7c0153 | 450 | struct nfs_commit_data *wdata = data; |
e0c2b380 | 451 | |
0b7c0153 FI |
452 | if (nfs41_setup_sequence(wdata->ds_clp->cl_session, |
453 | &wdata->args.seq_args, &wdata->res.seq_res, | |
454 | task)) | |
455 | return; | |
456 | ||
457 | rpc_call_start(task); | |
458 | } | |
459 | ||
460 | static void filelayout_write_commit_done(struct rpc_task *task, void *data) | |
461 | { | |
462 | struct nfs_commit_data *wdata = data; | |
463 | ||
464 | /* Note this may cause RPC to be resent */ | |
465 | wdata->mds_ops->rpc_call_done(task, data); | |
466 | } | |
467 | ||
468 | static void filelayout_commit_count_stats(struct rpc_task *task, void *data) | |
469 | { | |
470 | struct nfs_commit_data *cdata = data; | |
471 | ||
472 | rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics); | |
473 | } | |
474 | ||
475 | static void filelayout_commit_release(void *calldata) | |
476 | { | |
477 | struct nfs_commit_data *data = calldata; | |
478 | ||
f453a54a | 479 | data->completion_ops->completion(data); |
9369a431 | 480 | pnfs_put_lseg(data->lseg); |
3a7936c3 | 481 | nfs_put_client(data->ds_clp); |
0b7c0153 | 482 | nfs_commitdata_release(data); |
e0c2b380 FI |
483 | } |
484 | ||
17280175 | 485 | static const struct rpc_call_ops filelayout_read_call_ops = { |
dc70d7b3 AA |
486 | .rpc_call_prepare = filelayout_read_prepare, |
487 | .rpc_call_done = filelayout_read_call_done, | |
0a702195 | 488 | .rpc_count_stats = filelayout_read_count_stats, |
dc70d7b3 AA |
489 | .rpc_release = filelayout_read_release, |
490 | }; | |
491 | ||
17280175 | 492 | static const struct rpc_call_ops filelayout_write_call_ops = { |
a69aef14 FI |
493 | .rpc_call_prepare = filelayout_write_prepare, |
494 | .rpc_call_done = filelayout_write_call_done, | |
0a702195 | 495 | .rpc_count_stats = filelayout_write_count_stats, |
a69aef14 FI |
496 | .rpc_release = filelayout_write_release, |
497 | }; | |
498 | ||
17280175 | 499 | static const struct rpc_call_ops filelayout_commit_call_ops = { |
0b7c0153 FI |
500 | .rpc_call_prepare = filelayout_commit_prepare, |
501 | .rpc_call_done = filelayout_write_commit_done, | |
502 | .rpc_count_stats = filelayout_commit_count_stats, | |
e0c2b380 FI |
503 | .rpc_release = filelayout_commit_release, |
504 | }; | |
505 | ||
dc70d7b3 AA |
506 | static enum pnfs_try_status |
507 | filelayout_read_pagelist(struct nfs_read_data *data) | |
508 | { | |
cd841605 FI |
509 | struct nfs_pgio_header *hdr = data->header; |
510 | struct pnfs_layout_segment *lseg = hdr->lseg; | |
dc70d7b3 AA |
511 | struct nfs4_pnfs_ds *ds; |
512 | loff_t offset = data->args.offset; | |
513 | u32 j, idx; | |
514 | struct nfs_fh *fh; | |
dc70d7b3 AA |
515 | |
516 | dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n", | |
cd841605 | 517 | __func__, hdr->inode->i_ino, |
dc70d7b3 AA |
518 | data->args.pgbase, (size_t)data->args.count, offset); |
519 | ||
520 | /* Retrieve the correct rpc_client for the byte range */ | |
521 | j = nfs4_fl_calc_j_index(lseg, offset); | |
522 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
523 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 524 | if (!ds) |
dc70d7b3 | 525 | return PNFS_NOT_ATTEMPTED; |
3a7936c3 AA |
526 | dprintk("%s USE DS: %s cl_count %d\n", __func__, |
527 | ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count)); | |
dc70d7b3 AA |
528 | |
529 | /* No multipath support. Use first DS */ | |
3a7936c3 | 530 | atomic_inc(&ds->ds_clp->cl_count); |
dc70d7b3 AA |
531 | data->ds_clp = ds->ds_clp; |
532 | fh = nfs4_fl_select_ds_fh(lseg, j); | |
533 | if (fh) | |
534 | data->args.fh = fh; | |
535 | ||
536 | data->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
537 | data->mds_offset = offset; | |
538 | ||
539 | /* Perform an asynchronous read to ds */ | |
bc5a89b3 | 540 | nfs_initiate_read(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 | 541 | &filelayout_read_call_ops, RPC_TASK_SOFTCONN); |
dc70d7b3 AA |
542 | return PNFS_ATTEMPTED; |
543 | } | |
544 | ||
a69aef14 | 545 | /* Perform async writes. */ |
0382b744 AA |
546 | static enum pnfs_try_status |
547 | filelayout_write_pagelist(struct nfs_write_data *data, int sync) | |
548 | { | |
cd841605 FI |
549 | struct nfs_pgio_header *hdr = data->header; |
550 | struct pnfs_layout_segment *lseg = hdr->lseg; | |
a69aef14 FI |
551 | struct nfs4_pnfs_ds *ds; |
552 | loff_t offset = data->args.offset; | |
553 | u32 j, idx; | |
554 | struct nfs_fh *fh; | |
a69aef14 FI |
555 | |
556 | /* Retrieve the correct rpc_client for the byte range */ | |
557 | j = nfs4_fl_calc_j_index(lseg, offset); | |
558 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
559 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 560 | if (!ds) |
a69aef14 | 561 | return PNFS_NOT_ATTEMPTED; |
3a7936c3 AA |
562 | dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n", |
563 | __func__, hdr->inode->i_ino, sync, (size_t) data->args.count, | |
564 | offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count)); | |
a69aef14 | 565 | |
a69aef14 | 566 | data->write_done_cb = filelayout_write_done_cb; |
3a7936c3 | 567 | atomic_inc(&ds->ds_clp->cl_count); |
a69aef14 FI |
568 | data->ds_clp = ds->ds_clp; |
569 | fh = nfs4_fl_select_ds_fh(lseg, j); | |
570 | if (fh) | |
571 | data->args.fh = fh; | |
572 | /* | |
573 | * Get the file offset on the dserver. Set the write offset to | |
574 | * this offset and save the original offset. | |
575 | */ | |
576 | data->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
a69aef14 FI |
577 | |
578 | /* Perform an asynchronous write */ | |
bc5a89b3 | 579 | nfs_initiate_write(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 AA |
580 | &filelayout_write_call_ops, sync, |
581 | RPC_TASK_SOFTCONN); | |
a69aef14 | 582 | return PNFS_ATTEMPTED; |
0382b744 AA |
583 | } |
584 | ||
16b374ca AA |
585 | /* |
586 | * filelayout_check_layout() | |
587 | * | |
588 | * Make sure layout segment parameters are sane WRT the device. | |
589 | * At this point no generic layer initialization of the lseg has occurred, | |
590 | * and nothing has been added to the layout_hdr cache. | |
591 | * | |
592 | */ | |
593 | static int | |
594 | filelayout_check_layout(struct pnfs_layout_hdr *lo, | |
595 | struct nfs4_filelayout_segment *fl, | |
596 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 TM |
597 | struct nfs4_deviceid *id, |
598 | gfp_t gfp_flags) | |
16b374ca | 599 | { |
a1eaecbc | 600 | struct nfs4_deviceid_node *d; |
16b374ca AA |
601 | struct nfs4_file_layout_dsaddr *dsaddr; |
602 | int status = -EINVAL; | |
b7edfaa1 | 603 | struct nfs_server *nfss = NFS_SERVER(lo->plh_inode); |
16b374ca AA |
604 | |
605 | dprintk("--> %s\n", __func__); | |
606 | ||
7c24d948 AA |
607 | /* FIXME: remove this check when layout segment support is added */ |
608 | if (lgr->range.offset != 0 || | |
609 | lgr->range.length != NFS4_MAX_UINT64) { | |
610 | dprintk("%s Only whole file layouts supported. Use MDS i/o\n", | |
611 | __func__); | |
612 | goto out; | |
613 | } | |
614 | ||
16b374ca | 615 | if (fl->pattern_offset > lgr->range.offset) { |
3b6445a6 | 616 | dprintk("%s pattern_offset %lld too large\n", |
16b374ca AA |
617 | __func__, fl->pattern_offset); |
618 | goto out; | |
619 | } | |
620 | ||
75247aff BH |
621 | if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) { |
622 | dprintk("%s Invalid stripe unit (%u)\n", | |
16b374ca AA |
623 | __func__, fl->stripe_unit); |
624 | goto out; | |
625 | } | |
626 | ||
627 | /* find and reference the deviceid */ | |
35c8bb54 BH |
628 | d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld, |
629 | NFS_SERVER(lo->plh_inode)->nfs_client, id); | |
a1eaecbc | 630 | if (d == NULL) { |
78e4e05c | 631 | dsaddr = filelayout_get_device_info(lo->plh_inode, id, gfp_flags); |
16b374ca AA |
632 | if (dsaddr == NULL) |
633 | goto out; | |
a1eaecbc BH |
634 | } else |
635 | dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node); | |
1dfed273 TM |
636 | /* Found deviceid is unavailable */ |
637 | if (filelayout_test_devid_unavailable(&dsaddr->id_node)) | |
c47abcf8 AA |
638 | goto out_put; |
639 | ||
16b374ca AA |
640 | fl->dsaddr = dsaddr; |
641 | ||
e414966b CL |
642 | if (fl->first_stripe_index >= dsaddr->stripe_count) { |
643 | dprintk("%s Bad first_stripe_index %u\n", | |
16b374ca AA |
644 | __func__, fl->first_stripe_index); |
645 | goto out_put; | |
646 | } | |
647 | ||
648 | if ((fl->stripe_type == STRIPE_SPARSE && | |
649 | fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) || | |
650 | (fl->stripe_type == STRIPE_DENSE && | |
651 | fl->num_fh != dsaddr->stripe_count)) { | |
652 | dprintk("%s num_fh %u not valid for given packing\n", | |
653 | __func__, fl->num_fh); | |
654 | goto out_put; | |
655 | } | |
656 | ||
657 | if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) { | |
658 | dprintk("%s Stripe unit (%u) not aligned with rsize %u " | |
659 | "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize, | |
660 | nfss->wsize); | |
661 | } | |
662 | ||
663 | status = 0; | |
664 | out: | |
665 | dprintk("--> %s returns %d\n", __func__, status); | |
666 | return status; | |
667 | out_put: | |
ea8eecdd | 668 | nfs4_fl_put_deviceid(dsaddr); |
16b374ca AA |
669 | goto out; |
670 | } | |
671 | ||
672 | static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl) | |
673 | { | |
674 | int i; | |
675 | ||
676 | for (i = 0; i < fl->num_fh; i++) { | |
677 | if (!fl->fh_array[i]) | |
678 | break; | |
679 | kfree(fl->fh_array[i]); | |
680 | } | |
681 | kfree(fl->fh_array); | |
682 | fl->fh_array = NULL; | |
683 | } | |
684 | ||
685 | static void | |
686 | _filelayout_free_lseg(struct nfs4_filelayout_segment *fl) | |
687 | { | |
688 | filelayout_free_fh_array(fl); | |
689 | kfree(fl); | |
690 | } | |
691 | ||
692 | static int | |
693 | filelayout_decode_layout(struct pnfs_layout_hdr *flo, | |
694 | struct nfs4_filelayout_segment *fl, | |
695 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 TM |
696 | struct nfs4_deviceid *id, |
697 | gfp_t gfp_flags) | |
16b374ca | 698 | { |
35124a09 | 699 | struct xdr_stream stream; |
f7da7a12 | 700 | struct xdr_buf buf; |
35124a09 WAA |
701 | struct page *scratch; |
702 | __be32 *p; | |
16b374ca AA |
703 | uint32_t nfl_util; |
704 | int i; | |
705 | ||
706 | dprintk("%s: set_layout_map Begin\n", __func__); | |
707 | ||
a75b9df9 | 708 | scratch = alloc_page(gfp_flags); |
35124a09 WAA |
709 | if (!scratch) |
710 | return -ENOMEM; | |
711 | ||
f7da7a12 | 712 | xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len); |
35124a09 WAA |
713 | xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE); |
714 | ||
715 | /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8), | |
716 | * num_fh (4) */ | |
717 | p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20); | |
718 | if (unlikely(!p)) | |
719 | goto out_err; | |
720 | ||
16b374ca AA |
721 | memcpy(id, p, sizeof(*id)); |
722 | p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE); | |
a1eaecbc | 723 | nfs4_print_deviceid(id); |
16b374ca AA |
724 | |
725 | nfl_util = be32_to_cpup(p++); | |
726 | if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS) | |
727 | fl->commit_through_mds = 1; | |
728 | if (nfl_util & NFL4_UFLG_DENSE) | |
729 | fl->stripe_type = STRIPE_DENSE; | |
730 | else | |
731 | fl->stripe_type = STRIPE_SPARSE; | |
732 | fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK; | |
733 | ||
734 | fl->first_stripe_index = be32_to_cpup(p++); | |
735 | p = xdr_decode_hyper(p, &fl->pattern_offset); | |
736 | fl->num_fh = be32_to_cpup(p++); | |
737 | ||
738 | dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n", | |
739 | __func__, nfl_util, fl->num_fh, fl->first_stripe_index, | |
740 | fl->pattern_offset); | |
741 | ||
cec765cf AA |
742 | /* Note that a zero value for num_fh is legal for STRIPE_SPARSE. |
743 | * Futher checking is done in filelayout_check_layout */ | |
e414966b | 744 | if (fl->num_fh > |
cec765cf | 745 | max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT)) |
35124a09 WAA |
746 | goto out_err; |
747 | ||
cec765cf | 748 | if (fl->num_fh > 0) { |
1813badd | 749 | fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]), |
cec765cf AA |
750 | gfp_flags); |
751 | if (!fl->fh_array) | |
752 | goto out_err; | |
753 | } | |
16b374ca AA |
754 | |
755 | for (i = 0; i < fl->num_fh; i++) { | |
756 | /* Do we want to use a mempool here? */ | |
a75b9df9 | 757 | fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags); |
35124a09 WAA |
758 | if (!fl->fh_array[i]) |
759 | goto out_err_free; | |
760 | ||
761 | p = xdr_inline_decode(&stream, 4); | |
762 | if (unlikely(!p)) | |
763 | goto out_err_free; | |
16b374ca AA |
764 | fl->fh_array[i]->size = be32_to_cpup(p++); |
765 | if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) { | |
f9fd2d9c | 766 | printk(KERN_ERR "NFS: Too big fh %d received %d\n", |
16b374ca | 767 | i, fl->fh_array[i]->size); |
35124a09 | 768 | goto out_err_free; |
16b374ca | 769 | } |
35124a09 WAA |
770 | |
771 | p = xdr_inline_decode(&stream, fl->fh_array[i]->size); | |
772 | if (unlikely(!p)) | |
773 | goto out_err_free; | |
16b374ca | 774 | memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size); |
16b374ca AA |
775 | dprintk("DEBUG: %s: fh len %d\n", __func__, |
776 | fl->fh_array[i]->size); | |
777 | } | |
778 | ||
35124a09 | 779 | __free_page(scratch); |
16b374ca | 780 | return 0; |
35124a09 WAA |
781 | |
782 | out_err_free: | |
783 | filelayout_free_fh_array(fl); | |
784 | out_err: | |
785 | __free_page(scratch); | |
786 | return -EIO; | |
16b374ca AA |
787 | } |
788 | ||
c879513e FI |
789 | static void |
790 | filelayout_free_lseg(struct pnfs_layout_segment *lseg) | |
791 | { | |
792 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
793 | ||
794 | dprintk("--> %s\n", __func__); | |
795 | nfs4_fl_put_deviceid(fl->dsaddr); | |
799ba8d5 FI |
796 | /* This assumes a single RW lseg */ |
797 | if (lseg->pls_range.iomode == IOMODE_RW) { | |
798 | struct nfs4_filelayout *flo; | |
799 | ||
800 | flo = FILELAYOUT_FROM_HDR(lseg->pls_layout); | |
801 | flo->commit_info.nbuckets = 0; | |
802 | kfree(flo->commit_info.buckets); | |
803 | flo->commit_info.buckets = NULL; | |
804 | } | |
c879513e FI |
805 | _filelayout_free_lseg(fl); |
806 | } | |
807 | ||
799ba8d5 FI |
808 | static int |
809 | filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg, | |
ea2cf228 | 810 | struct nfs_commit_info *cinfo, |
799ba8d5 FI |
811 | gfp_t gfp_flags) |
812 | { | |
813 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
ea2cf228 | 814 | struct pnfs_commit_bucket *buckets; |
799ba8d5 FI |
815 | int size; |
816 | ||
817 | if (fl->commit_through_mds) | |
818 | return 0; | |
ea2cf228 | 819 | if (cinfo->ds->nbuckets != 0) { |
799ba8d5 FI |
820 | /* This assumes there is only one IOMODE_RW lseg. What |
821 | * we really want to do is have a layout_hdr level | |
822 | * dictionary of <multipath_list4, fh> keys, each | |
823 | * associated with a struct list_head, populated by calls | |
824 | * to filelayout_write_pagelist(). | |
825 | * */ | |
826 | return 0; | |
827 | } | |
828 | ||
829 | size = (fl->stripe_type == STRIPE_SPARSE) ? | |
830 | fl->dsaddr->ds_num : fl->dsaddr->stripe_count; | |
831 | ||
ea2cf228 | 832 | buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket), |
799ba8d5 FI |
833 | gfp_flags); |
834 | if (!buckets) | |
835 | return -ENOMEM; | |
836 | else { | |
837 | int i; | |
838 | ||
ea2cf228 FI |
839 | spin_lock(cinfo->lock); |
840 | if (cinfo->ds->nbuckets != 0) | |
799ba8d5 FI |
841 | kfree(buckets); |
842 | else { | |
ea2cf228 FI |
843 | cinfo->ds->buckets = buckets; |
844 | cinfo->ds->nbuckets = size; | |
799ba8d5 FI |
845 | for (i = 0; i < size; i++) { |
846 | INIT_LIST_HEAD(&buckets[i].written); | |
847 | INIT_LIST_HEAD(&buckets[i].committing); | |
848 | } | |
849 | } | |
ea2cf228 | 850 | spin_unlock(cinfo->lock); |
799ba8d5 FI |
851 | return 0; |
852 | } | |
853 | } | |
854 | ||
16b374ca AA |
855 | static struct pnfs_layout_segment * |
856 | filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid, | |
a75b9df9 TM |
857 | struct nfs4_layoutget_res *lgr, |
858 | gfp_t gfp_flags) | |
16b374ca AA |
859 | { |
860 | struct nfs4_filelayout_segment *fl; | |
861 | int rc; | |
862 | struct nfs4_deviceid id; | |
863 | ||
864 | dprintk("--> %s\n", __func__); | |
a75b9df9 | 865 | fl = kzalloc(sizeof(*fl), gfp_flags); |
16b374ca AA |
866 | if (!fl) |
867 | return NULL; | |
868 | ||
a75b9df9 TM |
869 | rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags); |
870 | if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) { | |
16b374ca AA |
871 | _filelayout_free_lseg(fl); |
872 | return NULL; | |
873 | } | |
874 | return &fl->generic_hdr; | |
875 | } | |
876 | ||
94ad1c80 FI |
877 | /* |
878 | * filelayout_pg_test(). Called by nfs_can_coalesce_requests() | |
879 | * | |
18ad0a9f BH |
880 | * return true : coalesce page |
881 | * return false : don't coalesce page | |
94ad1c80 | 882 | */ |
1751c363 | 883 | static bool |
94ad1c80 FI |
884 | filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev, |
885 | struct nfs_page *req) | |
886 | { | |
887 | u64 p_stripe, r_stripe; | |
888 | u32 stripe_unit; | |
889 | ||
19345cb2 BH |
890 | if (!pnfs_generic_pg_test(pgio, prev, req) || |
891 | !nfs_generic_pg_test(pgio, prev, req)) | |
892 | return false; | |
89a58e32 | 893 | |
b5542849 FI |
894 | p_stripe = (u64)req_offset(prev); |
895 | r_stripe = (u64)req_offset(req); | |
94ad1c80 FI |
896 | stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit; |
897 | ||
898 | do_div(p_stripe, stripe_unit); | |
899 | do_div(r_stripe, stripe_unit); | |
900 | ||
901 | return (p_stripe == r_stripe); | |
902 | } | |
903 | ||
17280175 | 904 | static void |
7c24d948 AA |
905 | filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio, |
906 | struct nfs_page *req) | |
907 | { | |
bc5a89b3 | 908 | WARN_ON_ONCE(pgio->pg_lseg != NULL); |
7c24d948 | 909 | |
1825a0d0 FI |
910 | if (req->wb_offset != req->wb_pgbase) { |
911 | /* | |
912 | * Handling unaligned pages is difficult, because have to | |
913 | * somehow split a req in two in certain cases in the | |
914 | * pg.test code. Avoid this by just not using pnfs | |
915 | * in this case. | |
916 | */ | |
917 | nfs_pageio_reset_read_mds(pgio); | |
918 | return; | |
919 | } | |
7c24d948 AA |
920 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
921 | req->wb_context, | |
922 | 0, | |
923 | NFS4_MAX_UINT64, | |
924 | IOMODE_READ, | |
925 | GFP_KERNEL); | |
926 | /* If no lseg, fall back to read through mds */ | |
927 | if (pgio->pg_lseg == NULL) | |
1f945357 | 928 | nfs_pageio_reset_read_mds(pgio); |
7c24d948 AA |
929 | } |
930 | ||
17280175 | 931 | static void |
7c24d948 AA |
932 | filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio, |
933 | struct nfs_page *req) | |
934 | { | |
ea2cf228 | 935 | struct nfs_commit_info cinfo; |
799ba8d5 FI |
936 | int status; |
937 | ||
bc5a89b3 | 938 | WARN_ON_ONCE(pgio->pg_lseg != NULL); |
7c24d948 | 939 | |
1825a0d0 FI |
940 | if (req->wb_offset != req->wb_pgbase) |
941 | goto out_mds; | |
7c24d948 AA |
942 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
943 | req->wb_context, | |
944 | 0, | |
945 | NFS4_MAX_UINT64, | |
946 | IOMODE_RW, | |
947 | GFP_NOFS); | |
948 | /* If no lseg, fall back to write through mds */ | |
949 | if (pgio->pg_lseg == NULL) | |
799ba8d5 | 950 | goto out_mds; |
ea2cf228 FI |
951 | nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq); |
952 | status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS); | |
799ba8d5 | 953 | if (status < 0) { |
9369a431 | 954 | pnfs_put_lseg(pgio->pg_lseg); |
799ba8d5 FI |
955 | pgio->pg_lseg = NULL; |
956 | goto out_mds; | |
957 | } | |
958 | return; | |
959 | out_mds: | |
960 | nfs_pageio_reset_write_mds(pgio); | |
7c24d948 AA |
961 | } |
962 | ||
1751c363 | 963 | static const struct nfs_pageio_ops filelayout_pg_read_ops = { |
7c24d948 | 964 | .pg_init = filelayout_pg_init_read, |
1751c363 | 965 | .pg_test = filelayout_pg_test, |
493292dd | 966 | .pg_doio = pnfs_generic_pg_readpages, |
1751c363 TM |
967 | }; |
968 | ||
969 | static const struct nfs_pageio_ops filelayout_pg_write_ops = { | |
7c24d948 | 970 | .pg_init = filelayout_pg_init_write, |
1751c363 | 971 | .pg_test = filelayout_pg_test, |
dce81290 | 972 | .pg_doio = pnfs_generic_pg_writepages, |
1751c363 TM |
973 | }; |
974 | ||
e0c2b380 FI |
975 | static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j) |
976 | { | |
977 | if (fl->stripe_type == STRIPE_SPARSE) | |
978 | return nfs4_fl_calc_ds_index(&fl->generic_hdr, j); | |
979 | else | |
980 | return j; | |
981 | } | |
982 | ||
d6d6dc7c FI |
983 | /* The generic layer is about to remove the req from the commit list. |
984 | * If this will make the bucket empty, it will need to put the lseg reference. | |
d6d6dc7c | 985 | */ |
8dd37758 | 986 | static void |
ea2cf228 FI |
987 | filelayout_clear_request_commit(struct nfs_page *req, |
988 | struct nfs_commit_info *cinfo) | |
d6d6dc7c | 989 | { |
8dd37758 | 990 | struct pnfs_layout_segment *freeme = NULL; |
8dd37758 | 991 | |
ea2cf228 | 992 | spin_lock(cinfo->lock); |
8dd37758 TM |
993 | if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags)) |
994 | goto out; | |
ea2cf228 | 995 | cinfo->ds->nwritten--; |
d6d6dc7c | 996 | if (list_is_singular(&req->wb_list)) { |
ea2cf228 | 997 | struct pnfs_commit_bucket *bucket; |
d6d6dc7c | 998 | |
799ba8d5 | 999 | bucket = list_first_entry(&req->wb_list, |
ea2cf228 | 1000 | struct pnfs_commit_bucket, |
799ba8d5 FI |
1001 | written); |
1002 | freeme = bucket->wlseg; | |
1003 | bucket->wlseg = NULL; | |
d6d6dc7c | 1004 | } |
8dd37758 | 1005 | out: |
ea2cf228 FI |
1006 | nfs_request_remove_commit_list(req, cinfo); |
1007 | spin_unlock(cinfo->lock); | |
9369a431 | 1008 | pnfs_put_lseg(freeme); |
d6d6dc7c FI |
1009 | } |
1010 | ||
1011 | static struct list_head * | |
1012 | filelayout_choose_commit_list(struct nfs_page *req, | |
ea2cf228 FI |
1013 | struct pnfs_layout_segment *lseg, |
1014 | struct nfs_commit_info *cinfo) | |
e0c2b380 | 1015 | { |
e0c2b380 FI |
1016 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); |
1017 | u32 i, j; | |
1018 | struct list_head *list; | |
ea2cf228 | 1019 | struct pnfs_commit_bucket *buckets; |
e0c2b380 | 1020 | |
d6d6dc7c | 1021 | if (fl->commit_through_mds) |
ea2cf228 | 1022 | return &cinfo->mds->list; |
d6d6dc7c | 1023 | |
e0c2b380 FI |
1024 | /* Note that we are calling nfs4_fl_calc_j_index on each page |
1025 | * that ends up being committed to a data server. An attractive | |
1026 | * alternative is to add a field to nfs_write_data and nfs_page | |
1027 | * to store the value calculated in filelayout_write_pagelist | |
1028 | * and just use that here. | |
1029 | */ | |
b5542849 | 1030 | j = nfs4_fl_calc_j_index(lseg, req_offset(req)); |
e0c2b380 | 1031 | i = select_bucket_index(fl, j); |
ea2cf228 | 1032 | buckets = cinfo->ds->buckets; |
799ba8d5 | 1033 | list = &buckets[i].written; |
e0c2b380 | 1034 | if (list_empty(list)) { |
d6d6dc7c FI |
1035 | /* Non-empty buckets hold a reference on the lseg. That ref |
1036 | * is normally transferred to the COMMIT call and released | |
1037 | * there. It could also be released if the last req is pulled | |
1038 | * off due to a rewrite, in which case it will be done in | |
799ba8d5 | 1039 | * filelayout_clear_request_commit |
d6d6dc7c | 1040 | */ |
9369a431 | 1041 | buckets[i].wlseg = pnfs_get_lseg(lseg); |
e0c2b380 | 1042 | } |
8dd37758 | 1043 | set_bit(PG_COMMIT_TO_DS, &req->wb_flags); |
ea2cf228 | 1044 | cinfo->ds->nwritten++; |
e0c2b380 FI |
1045 | return list; |
1046 | } | |
1047 | ||
8dd37758 TM |
1048 | static void |
1049 | filelayout_mark_request_commit(struct nfs_page *req, | |
ea2cf228 FI |
1050 | struct pnfs_layout_segment *lseg, |
1051 | struct nfs_commit_info *cinfo) | |
8dd37758 TM |
1052 | { |
1053 | struct list_head *list; | |
1054 | ||
ea2cf228 FI |
1055 | list = filelayout_choose_commit_list(req, lseg, cinfo); |
1056 | nfs_request_add_commit_list(req, list, cinfo); | |
8dd37758 TM |
1057 | } |
1058 | ||
e0c2b380 FI |
1059 | static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i) |
1060 | { | |
1061 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1062 | ||
1063 | if (flseg->stripe_type == STRIPE_SPARSE) | |
1064 | return i; | |
1065 | else | |
1066 | return nfs4_fl_calc_ds_index(lseg, i); | |
1067 | } | |
1068 | ||
1069 | static struct nfs_fh * | |
1070 | select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i) | |
1071 | { | |
1072 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1073 | ||
1074 | if (flseg->stripe_type == STRIPE_SPARSE) { | |
1075 | if (flseg->num_fh == 1) | |
1076 | i = 0; | |
1077 | else if (flseg->num_fh == 0) | |
1078 | /* Use the MDS OPEN fh set in nfs_read_rpcsetup */ | |
1079 | return NULL; | |
1080 | } | |
1081 | return flseg->fh_array[i]; | |
1082 | } | |
1083 | ||
0b7c0153 | 1084 | static int filelayout_initiate_commit(struct nfs_commit_data *data, int how) |
e0c2b380 FI |
1085 | { |
1086 | struct pnfs_layout_segment *lseg = data->lseg; | |
1087 | struct nfs4_pnfs_ds *ds; | |
1088 | u32 idx; | |
1089 | struct nfs_fh *fh; | |
1090 | ||
1091 | idx = calc_ds_index_from_commit(lseg, data->ds_commit_index); | |
1092 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
1093 | if (!ds) { | |
e0c2b380 | 1094 | prepare_to_resend_writes(data); |
8dd37758 | 1095 | filelayout_commit_release(data); |
e0c2b380 FI |
1096 | return -EAGAIN; |
1097 | } | |
3a7936c3 AA |
1098 | dprintk("%s ino %lu, how %d cl_count %d\n", __func__, |
1099 | data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count)); | |
0b7c0153 | 1100 | data->commit_done_cb = filelayout_commit_done_cb; |
3a7936c3 | 1101 | atomic_inc(&ds->ds_clp->cl_count); |
e0c2b380 FI |
1102 | data->ds_clp = ds->ds_clp; |
1103 | fh = select_ds_fh_from_commit(lseg, data->ds_commit_index); | |
1104 | if (fh) | |
1105 | data->args.fh = fh; | |
0b7c0153 | 1106 | return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 AA |
1107 | &filelayout_commit_call_ops, how, |
1108 | RPC_TASK_SOFTCONN); | |
e0c2b380 FI |
1109 | } |
1110 | ||
8dd37758 | 1111 | static int |
1763da12 FI |
1112 | transfer_commit_list(struct list_head *src, struct list_head *dst, |
1113 | struct nfs_commit_info *cinfo, int max) | |
8dd37758 | 1114 | { |
8dd37758 TM |
1115 | struct nfs_page *req, *tmp; |
1116 | int ret = 0; | |
1117 | ||
1118 | list_for_each_entry_safe(req, tmp, src, wb_list) { | |
1119 | if (!nfs_lock_request(req)) | |
1120 | continue; | |
53b8ee34 | 1121 | kref_get(&req->wb_kref); |
ea2cf228 | 1122 | if (cond_resched_lock(cinfo->lock)) |
3b3be88d | 1123 | list_safe_reset_next(req, tmp, wb_list); |
ea2cf228 | 1124 | nfs_request_remove_commit_list(req, cinfo); |
8dd37758 TM |
1125 | clear_bit(PG_COMMIT_TO_DS, &req->wb_flags); |
1126 | nfs_list_add_request(req, dst); | |
1127 | ret++; | |
1763da12 | 1128 | if ((ret == max) && !cinfo->dreq) |
8dd37758 TM |
1129 | break; |
1130 | } | |
1763da12 FI |
1131 | return ret; |
1132 | } | |
1133 | ||
1134 | static int | |
1135 | filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket, | |
1136 | struct nfs_commit_info *cinfo, | |
1137 | int max) | |
1138 | { | |
1139 | struct list_head *src = &bucket->written; | |
1140 | struct list_head *dst = &bucket->committing; | |
1141 | int ret; | |
1142 | ||
1143 | ret = transfer_commit_list(src, dst, cinfo, max); | |
799ba8d5 | 1144 | if (ret) { |
ea2cf228 FI |
1145 | cinfo->ds->nwritten -= ret; |
1146 | cinfo->ds->ncommitting += ret; | |
799ba8d5 FI |
1147 | bucket->clseg = bucket->wlseg; |
1148 | if (list_empty(src)) | |
1149 | bucket->wlseg = NULL; | |
1150 | else | |
9369a431 | 1151 | pnfs_get_lseg(bucket->clseg); |
799ba8d5 | 1152 | } |
8dd37758 TM |
1153 | return ret; |
1154 | } | |
1155 | ||
d6d6dc7c | 1156 | /* Move reqs from written to committing lists, returning count of number moved. |
ea2cf228 | 1157 | * Note called with cinfo->lock held. |
d6d6dc7c | 1158 | */ |
ea2cf228 FI |
1159 | static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo, |
1160 | int max) | |
d6d6dc7c | 1161 | { |
d6d6dc7c FI |
1162 | int i, rv = 0, cnt; |
1163 | ||
ea2cf228 FI |
1164 | for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) { |
1165 | cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i], | |
1166 | cinfo, max); | |
d6d6dc7c FI |
1167 | max -= cnt; |
1168 | rv += cnt; | |
1169 | } | |
d6d6dc7c FI |
1170 | return rv; |
1171 | } | |
1172 | ||
1763da12 FI |
1173 | /* Pull everything off the committing lists and dump into @dst */ |
1174 | static void filelayout_recover_commit_reqs(struct list_head *dst, | |
1175 | struct nfs_commit_info *cinfo) | |
1176 | { | |
1177 | struct pnfs_commit_bucket *b; | |
1178 | int i; | |
1179 | ||
1180 | /* NOTE cinfo->lock is NOT held, relying on fact that this is | |
1181 | * only called on single thread per dreq. | |
9369a431 | 1182 | * Can't take the lock because need to do pnfs_put_lseg |
1763da12 FI |
1183 | */ |
1184 | for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) { | |
1185 | if (transfer_commit_list(&b->written, dst, cinfo, 0)) { | |
9369a431 | 1186 | pnfs_put_lseg(b->wlseg); |
1763da12 FI |
1187 | b->wlseg = NULL; |
1188 | } | |
1189 | } | |
1190 | cinfo->ds->nwritten = 0; | |
1191 | } | |
1192 | ||
9390f425 | 1193 | static unsigned int |
ea2cf228 | 1194 | alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list) |
e0c2b380 | 1195 | { |
ea2cf228 FI |
1196 | struct pnfs_ds_commit_info *fl_cinfo; |
1197 | struct pnfs_commit_bucket *bucket; | |
0b7c0153 | 1198 | struct nfs_commit_data *data; |
e0c2b380 | 1199 | int i, j; |
9390f425 | 1200 | unsigned int nreq = 0; |
e0c2b380 | 1201 | |
ea2cf228 | 1202 | fl_cinfo = cinfo->ds; |
799ba8d5 FI |
1203 | bucket = fl_cinfo->buckets; |
1204 | for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) { | |
1205 | if (list_empty(&bucket->committing)) | |
e0c2b380 FI |
1206 | continue; |
1207 | data = nfs_commitdata_alloc(); | |
1208 | if (!data) | |
9390f425 | 1209 | break; |
e0c2b380 | 1210 | data->ds_commit_index = i; |
799ba8d5 FI |
1211 | data->lseg = bucket->clseg; |
1212 | bucket->clseg = NULL; | |
e0c2b380 | 1213 | list_add(&data->pages, list); |
9390f425 | 1214 | nreq++; |
e0c2b380 | 1215 | } |
e0c2b380 | 1216 | |
9390f425 | 1217 | /* Clean up on error */ |
799ba8d5 FI |
1218 | for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) { |
1219 | if (list_empty(&bucket->committing)) | |
e0c2b380 | 1220 | continue; |
ea2cf228 | 1221 | nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo); |
9369a431 | 1222 | pnfs_put_lseg(bucket->clseg); |
799ba8d5 | 1223 | bucket->clseg = NULL; |
e0c2b380 | 1224 | } |
e0c2b380 | 1225 | /* Caller will clean up entries put on list */ |
9390f425 | 1226 | return nreq; |
e0c2b380 FI |
1227 | } |
1228 | ||
1229 | /* This follows nfs_commit_list pretty closely */ | |
1230 | static int | |
1231 | filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages, | |
ea2cf228 | 1232 | int how, struct nfs_commit_info *cinfo) |
e0c2b380 | 1233 | { |
0b7c0153 | 1234 | struct nfs_commit_data *data, *tmp; |
e0c2b380 | 1235 | LIST_HEAD(list); |
9390f425 | 1236 | unsigned int nreq = 0; |
e0c2b380 FI |
1237 | |
1238 | if (!list_empty(mds_pages)) { | |
1239 | data = nfs_commitdata_alloc(); | |
9390f425 TM |
1240 | if (data != NULL) { |
1241 | data->lseg = NULL; | |
1242 | list_add(&data->pages, &list); | |
1243 | nreq++; | |
1244 | } else | |
ea2cf228 | 1245 | nfs_retry_commit(mds_pages, NULL, cinfo); |
e0c2b380 FI |
1246 | } |
1247 | ||
ea2cf228 | 1248 | nreq += alloc_ds_commits(cinfo, &list); |
9390f425 TM |
1249 | |
1250 | if (nreq == 0) { | |
f453a54a | 1251 | cinfo->completion_ops->error_cleanup(NFS_I(inode)); |
9390f425 TM |
1252 | goto out; |
1253 | } | |
1254 | ||
ea2cf228 | 1255 | atomic_add(nreq, &cinfo->mds->rpcs_out); |
e0c2b380 FI |
1256 | |
1257 | list_for_each_entry_safe(data, tmp, &list, pages) { | |
1258 | list_del_init(&data->pages); | |
e0c2b380 | 1259 | if (!data->lseg) { |
f453a54a | 1260 | nfs_init_commit(data, mds_pages, NULL, cinfo); |
0b7c0153 | 1261 | nfs_initiate_commit(NFS_CLIENT(inode), data, |
9f0ec176 | 1262 | data->mds_ops, how, 0); |
e0c2b380 | 1263 | } else { |
ea2cf228 | 1264 | struct pnfs_commit_bucket *buckets; |
799ba8d5 | 1265 | |
ea2cf228 | 1266 | buckets = cinfo->ds->buckets; |
f453a54a | 1267 | nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo); |
e0c2b380 FI |
1268 | filelayout_initiate_commit(data, how); |
1269 | } | |
1270 | } | |
9390f425 | 1271 | out: |
ea2cf228 | 1272 | cinfo->ds->ncommitting = 0; |
9390f425 | 1273 | return PNFS_ATTEMPTED; |
e0c2b380 FI |
1274 | } |
1275 | ||
1775bc34 BH |
1276 | static void |
1277 | filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d) | |
1278 | { | |
1279 | nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node)); | |
1280 | } | |
1281 | ||
799ba8d5 FI |
1282 | static struct pnfs_layout_hdr * |
1283 | filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags) | |
1284 | { | |
1285 | struct nfs4_filelayout *flo; | |
1286 | ||
1287 | flo = kzalloc(sizeof(*flo), gfp_flags); | |
1288 | return &flo->generic_hdr; | |
1289 | } | |
1290 | ||
1291 | static void | |
1292 | filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo) | |
1293 | { | |
1294 | kfree(FILELAYOUT_FROM_HDR(lo)); | |
1295 | } | |
1296 | ||
ea2cf228 FI |
1297 | static struct pnfs_ds_commit_info * |
1298 | filelayout_get_ds_info(struct inode *inode) | |
1299 | { | |
df011748 FI |
1300 | struct pnfs_layout_hdr *layout = NFS_I(inode)->layout; |
1301 | ||
1302 | if (layout == NULL) | |
1303 | return NULL; | |
1304 | else | |
1305 | return &FILELAYOUT_FROM_HDR(layout)->commit_info; | |
ea2cf228 FI |
1306 | } |
1307 | ||
7ab672ce | 1308 | static struct pnfs_layoutdriver_type filelayout_type = { |
ea8eecdd CH |
1309 | .id = LAYOUT_NFSV4_1_FILES, |
1310 | .name = "LAYOUT_NFSV4_1_FILES", | |
1311 | .owner = THIS_MODULE, | |
799ba8d5 FI |
1312 | .alloc_layout_hdr = filelayout_alloc_layout_hdr, |
1313 | .free_layout_hdr = filelayout_free_layout_hdr, | |
ea8eecdd CH |
1314 | .alloc_lseg = filelayout_alloc_lseg, |
1315 | .free_lseg = filelayout_free_lseg, | |
1751c363 TM |
1316 | .pg_read_ops = &filelayout_pg_read_ops, |
1317 | .pg_write_ops = &filelayout_pg_write_ops, | |
ea2cf228 | 1318 | .get_ds_info = &filelayout_get_ds_info, |
8dd37758 TM |
1319 | .mark_request_commit = filelayout_mark_request_commit, |
1320 | .clear_request_commit = filelayout_clear_request_commit, | |
d6d6dc7c | 1321 | .scan_commit_lists = filelayout_scan_commit_lists, |
1763da12 | 1322 | .recover_commit_reqs = filelayout_recover_commit_reqs, |
e0c2b380 | 1323 | .commit_pagelist = filelayout_commit_pagelist, |
dc70d7b3 | 1324 | .read_pagelist = filelayout_read_pagelist, |
0382b744 | 1325 | .write_pagelist = filelayout_write_pagelist, |
1775bc34 | 1326 | .free_deviceid_node = filelayout_free_deveiceid_node, |
7ab672ce DH |
1327 | }; |
1328 | ||
1329 | static int __init nfs4filelayout_init(void) | |
1330 | { | |
1331 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n", | |
1332 | __func__); | |
1333 | return pnfs_register_layoutdriver(&filelayout_type); | |
1334 | } | |
1335 | ||
1336 | static void __exit nfs4filelayout_exit(void) | |
1337 | { | |
1338 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n", | |
1339 | __func__); | |
1340 | pnfs_unregister_layoutdriver(&filelayout_type); | |
1341 | } | |
1342 | ||
f85ef69c BF |
1343 | MODULE_ALIAS("nfs-layouttype4-1"); |
1344 | ||
7ab672ce DH |
1345 | module_init(nfs4filelayout_init); |
1346 | module_exit(nfs4filelayout_exit); |