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 | ||
76e697ba | 38 | #include "nfs4session.h" |
16b374ca | 39 | #include "internal.h" |
9cb81968 | 40 | #include "delegation.h" |
16b374ca | 41 | #include "nfs4filelayout.h" |
7ab672ce DH |
42 | |
43 | #define NFSDBG_FACILITY NFSDBG_PNFS_LD | |
44 | ||
45 | MODULE_LICENSE("GPL"); | |
46 | MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>"); | |
47 | MODULE_DESCRIPTION("The NFSv4 file layout driver"); | |
48 | ||
cbdabc7f AA |
49 | #define FILELAYOUT_POLL_RETRY_MAX (15*HZ) |
50 | ||
cfe7f412 FI |
51 | static loff_t |
52 | filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg, | |
53 | loff_t offset) | |
54 | { | |
55 | u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count; | |
3476f114 CM |
56 | u64 stripe_no; |
57 | u32 rem; | |
cfe7f412 FI |
58 | |
59 | offset -= flseg->pattern_offset; | |
3476f114 CM |
60 | stripe_no = div_u64(offset, stripe_width); |
61 | div_u64_rem(offset, flseg->stripe_unit, &rem); | |
cfe7f412 | 62 | |
3476f114 | 63 | return stripe_no * flseg->stripe_unit + rem; |
cfe7f412 FI |
64 | } |
65 | ||
66 | /* This function is used by the layout driver to calculate the | |
67 | * offset of the file on the dserver based on whether the | |
68 | * layout type is STRIPE_DENSE or STRIPE_SPARSE | |
69 | */ | |
70 | static loff_t | |
71 | filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset) | |
72 | { | |
73 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
74 | ||
75 | switch (flseg->stripe_type) { | |
76 | case STRIPE_SPARSE: | |
77 | return offset; | |
78 | ||
79 | case STRIPE_DENSE: | |
80 | return filelayout_get_dense_offset(flseg, offset); | |
81 | } | |
82 | ||
83 | BUG(); | |
84 | } | |
85 | ||
e7dd79af AA |
86 | static void filelayout_reset_write(struct nfs_write_data *data) |
87 | { | |
88 | struct nfs_pgio_header *hdr = data->header; | |
e7dd79af AA |
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, | |
497826af BS |
95 | hdr->inode->i_sb->s_id, |
96 | (long long)NFS_FILEID(hdr->inode), | |
e7dd79af AA |
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, | |
78f33277 BH |
102 | hdr->completion_ops, |
103 | hdr->dreq); | |
e7dd79af AA |
104 | } |
105 | } | |
106 | ||
107 | static void filelayout_reset_read(struct nfs_read_data *data) | |
108 | { | |
109 | struct nfs_pgio_header *hdr = data->header; | |
e7dd79af AA |
110 | struct rpc_task *task = &data->task; |
111 | ||
112 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { | |
113 | dprintk("%s Reset task %5u for i/o through MDS " | |
114 | "(req %s/%lld, %u bytes @ offset %llu)\n", __func__, | |
115 | data->task.tk_pid, | |
497826af BS |
116 | hdr->inode->i_sb->s_id, |
117 | (long long)NFS_FILEID(hdr->inode), | |
e7dd79af AA |
118 | data->args.count, |
119 | (unsigned long long)data->args.offset); | |
120 | ||
121 | task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode, | |
122 | &hdr->pages, | |
78f33277 BH |
123 | hdr->completion_ops, |
124 | hdr->dreq); | |
e7dd79af AA |
125 | } |
126 | } | |
127 | ||
d527e5c1 TM |
128 | static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo) |
129 | { | |
130 | if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) | |
131 | return; | |
d527e5c1 TM |
132 | pnfs_return_layout(inode); |
133 | } | |
134 | ||
cbdabc7f AA |
135 | static int filelayout_async_handle_error(struct rpc_task *task, |
136 | struct nfs4_state *state, | |
137 | struct nfs_client *clp, | |
e7dd79af | 138 | struct pnfs_layout_segment *lseg) |
cbdabc7f | 139 | { |
d527e5c1 TM |
140 | struct pnfs_layout_hdr *lo = lseg->pls_layout; |
141 | struct inode *inode = lo->plh_inode; | |
e7dd79af AA |
142 | struct nfs_server *mds_server = NFS_SERVER(inode); |
143 | struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg); | |
9cb81968 | 144 | struct nfs_client *mds_client = mds_server->nfs_client; |
671fb896 | 145 | struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table; |
9cb81968 | 146 | |
cbdabc7f AA |
147 | if (task->tk_status >= 0) |
148 | return 0; | |
cbdabc7f AA |
149 | |
150 | switch (task->tk_status) { | |
9cb81968 AA |
151 | /* MDS state errors */ |
152 | case -NFS4ERR_DELEG_REVOKED: | |
153 | case -NFS4ERR_ADMIN_REVOKED: | |
154 | case -NFS4ERR_BAD_STATEID: | |
e7dd79af AA |
155 | if (state == NULL) |
156 | break; | |
2dc31756 | 157 | nfs_remove_bad_delegation(state->inode); |
9cb81968 | 158 | case -NFS4ERR_OPENMODE: |
e7dd79af AA |
159 | if (state == NULL) |
160 | break; | |
5d422301 TM |
161 | if (nfs4_schedule_stateid_recovery(mds_server, state) < 0) |
162 | goto out_bad_stateid; | |
9cb81968 AA |
163 | goto wait_on_recovery; |
164 | case -NFS4ERR_EXPIRED: | |
5d422301 TM |
165 | if (state != NULL) { |
166 | if (nfs4_schedule_stateid_recovery(mds_server, state) < 0) | |
167 | goto out_bad_stateid; | |
168 | } | |
9cb81968 AA |
169 | nfs4_schedule_lease_recovery(mds_client); |
170 | goto wait_on_recovery; | |
171 | /* DS session errors */ | |
cbdabc7f AA |
172 | case -NFS4ERR_BADSESSION: |
173 | case -NFS4ERR_BADSLOT: | |
174 | case -NFS4ERR_BAD_HIGH_SLOT: | |
175 | case -NFS4ERR_DEADSESSION: | |
176 | case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION: | |
177 | case -NFS4ERR_SEQ_FALSE_RETRY: | |
178 | case -NFS4ERR_SEQ_MISORDERED: | |
179 | dprintk("%s ERROR %d, Reset session. Exchangeid " | |
180 | "flags 0x%x\n", __func__, task->tk_status, | |
181 | clp->cl_exchange_flags); | |
9f594791 | 182 | nfs4_schedule_session_recovery(clp->cl_session, task->tk_status); |
cbdabc7f AA |
183 | break; |
184 | case -NFS4ERR_DELAY: | |
185 | case -NFS4ERR_GRACE: | |
cbdabc7f AA |
186 | rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX); |
187 | break; | |
a8a4ae3a AA |
188 | case -NFS4ERR_RETRY_UNCACHED_REP: |
189 | break; | |
041245c8 AA |
190 | /* Invalidate Layout errors */ |
191 | case -NFS4ERR_PNFS_NO_LAYOUT: | |
192 | case -ESTALE: /* mapped NFS4ERR_STALE */ | |
193 | case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */ | |
194 | case -EISDIR: /* mapped NFS4ERR_ISDIR */ | |
195 | case -NFS4ERR_FHEXPIRED: | |
196 | case -NFS4ERR_WRONG_TYPE: | |
197 | dprintk("%s Invalid layout error %d\n", __func__, | |
198 | task->tk_status); | |
199 | /* | |
200 | * Destroy layout so new i/o will get a new layout. | |
201 | * Layout will not be destroyed until all current lseg | |
202 | * references are put. Mark layout as invalid to resend failed | |
203 | * i/o and all i/o waiting on the slot table to the MDS until | |
204 | * layout is destroyed and a new valid layout is obtained. | |
205 | */ | |
d42e7873 | 206 | pnfs_destroy_layout(NFS_I(inode)); |
041245c8 AA |
207 | rpc_wake_up(&tbl->slot_tbl_waitq); |
208 | goto reset; | |
e7dd79af AA |
209 | /* RPC connection errors */ |
210 | case -ECONNREFUSED: | |
211 | case -EHOSTDOWN: | |
212 | case -EHOSTUNREACH: | |
213 | case -ENETUNREACH: | |
214 | case -EIO: | |
215 | case -ETIMEDOUT: | |
216 | case -EPIPE: | |
217 | dprintk("%s DS connection error %d\n", __func__, | |
218 | task->tk_status); | |
1dfed273 | 219 | nfs4_mark_deviceid_unavailable(devid); |
d527e5c1 | 220 | set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags); |
671fb896 | 221 | rpc_wake_up(&tbl->slot_tbl_waitq); |
e7dd79af | 222 | /* fall through */ |
cbdabc7f | 223 | default: |
041245c8 | 224 | reset: |
e7dd79af | 225 | dprintk("%s Retry through MDS. Error %d\n", __func__, |
cbdabc7f | 226 | task->tk_status); |
e7dd79af | 227 | return -NFS4ERR_RESET_TO_MDS; |
cbdabc7f | 228 | } |
9cb81968 | 229 | out: |
cbdabc7f AA |
230 | task->tk_status = 0; |
231 | return -EAGAIN; | |
5d422301 TM |
232 | out_bad_stateid: |
233 | task->tk_status = -EIO; | |
234 | return 0; | |
9cb81968 AA |
235 | wait_on_recovery: |
236 | rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL); | |
237 | if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0) | |
238 | rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task); | |
239 | goto out; | |
cbdabc7f AA |
240 | } |
241 | ||
242 | /* NFS_PROTO call done callback routines */ | |
243 | ||
244 | static int filelayout_read_done_cb(struct rpc_task *task, | |
245 | struct nfs_read_data *data) | |
246 | { | |
e7dd79af AA |
247 | struct nfs_pgio_header *hdr = data->header; |
248 | int err; | |
cbdabc7f | 249 | |
e7dd79af AA |
250 | err = filelayout_async_handle_error(task, data->args.context->state, |
251 | data->ds_clp, hdr->lseg); | |
cbdabc7f | 252 | |
e7dd79af AA |
253 | switch (err) { |
254 | case -NFS4ERR_RESET_TO_MDS: | |
255 | filelayout_reset_read(data); | |
256 | return task->tk_status; | |
257 | case -EAGAIN: | |
d00c5d43 | 258 | rpc_restart_call_prepare(task); |
cbdabc7f AA |
259 | return -EAGAIN; |
260 | } | |
261 | ||
262 | return 0; | |
263 | } | |
264 | ||
863a3c6c AA |
265 | /* |
266 | * We reference the rpc_cred of the first WRITE that triggers the need for | |
267 | * a LAYOUTCOMMIT, and use it to send the layoutcommit compound. | |
268 | * rfc5661 is not clear about which credential should be used. | |
269 | */ | |
270 | static void | |
271 | filelayout_set_layoutcommit(struct nfs_write_data *wdata) | |
272 | { | |
cd841605 FI |
273 | struct nfs_pgio_header *hdr = wdata->header; |
274 | ||
275 | if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds || | |
863a3c6c AA |
276 | wdata->res.verf->committed == NFS_FILE_SYNC) |
277 | return; | |
278 | ||
279 | pnfs_set_layoutcommit(wdata); | |
cd841605 FI |
280 | dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino, |
281 | (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb); | |
863a3c6c AA |
282 | } |
283 | ||
1dfed273 TM |
284 | bool |
285 | filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node) | |
286 | { | |
287 | return filelayout_test_devid_invalid(node) || | |
288 | nfs4_test_deviceid_unavailable(node); | |
289 | } | |
290 | ||
291 | static bool | |
292 | filelayout_reset_to_mds(struct pnfs_layout_segment *lseg) | |
293 | { | |
294 | struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg); | |
295 | ||
8006bfba | 296 | return filelayout_test_devid_unavailable(node); |
1dfed273 TM |
297 | } |
298 | ||
dc70d7b3 AA |
299 | /* |
300 | * Call ops for the async read/write cases | |
301 | * In the case of dense layouts, the offset needs to be reset to its | |
302 | * original value. | |
303 | */ | |
304 | static void filelayout_read_prepare(struct rpc_task *task, void *data) | |
305 | { | |
cd12ae32 | 306 | struct nfs_read_data *rdata = data; |
dc70d7b3 | 307 | |
c58c8441 TM |
308 | if (unlikely(test_bit(NFS_CONTEXT_BAD, &rdata->args.context->flags))) { |
309 | rpc_exit(task, -EIO); | |
310 | return; | |
311 | } | |
041245c8 | 312 | if (filelayout_reset_to_mds(rdata->header->lseg)) { |
0ad2f378 AA |
313 | dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid); |
314 | filelayout_reset_read(rdata); | |
315 | rpc_exit(task, 0); | |
316 | return; | |
317 | } | |
cbdabc7f AA |
318 | rdata->read_done_cb = filelayout_read_done_cb; |
319 | ||
9b206149 | 320 | if (nfs41_setup_sequence(rdata->ds_clp->cl_session, |
d9afbd1b TM |
321 | &rdata->args.seq_args, |
322 | &rdata->res.seq_res, | |
9b206149 TM |
323 | task)) |
324 | return; | |
325 | nfs4_set_rw_stateid(&rdata->args.stateid, rdata->args.context, | |
326 | rdata->args.lock_context, FMODE_READ); | |
dc70d7b3 AA |
327 | } |
328 | ||
329 | static void filelayout_read_call_done(struct rpc_task *task, void *data) | |
330 | { | |
cd12ae32 | 331 | struct nfs_read_data *rdata = data; |
dc70d7b3 AA |
332 | |
333 | dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status); | |
334 | ||
bd4aeffb AA |
335 | if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) && |
336 | task->tk_status == 0) | |
0ad2f378 AA |
337 | return; |
338 | ||
dc70d7b3 | 339 | /* Note this may cause RPC to be resent */ |
cd841605 | 340 | rdata->header->mds_ops->rpc_call_done(task, data); |
dc70d7b3 AA |
341 | } |
342 | ||
0a702195 WAA |
343 | static void filelayout_read_count_stats(struct rpc_task *task, void *data) |
344 | { | |
cd12ae32 | 345 | struct nfs_read_data *rdata = data; |
0a702195 | 346 | |
cd841605 | 347 | rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics); |
0a702195 WAA |
348 | } |
349 | ||
dc70d7b3 AA |
350 | static void filelayout_read_release(void *data) |
351 | { | |
cd12ae32 | 352 | struct nfs_read_data *rdata = data; |
d527e5c1 | 353 | struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout; |
dc70d7b3 | 354 | |
d527e5c1 | 355 | filelayout_fenceme(lo->plh_inode, lo); |
996074cb | 356 | nfs_put_client(rdata->ds_clp); |
cd841605 | 357 | rdata->header->mds_ops->rpc_release(data); |
dc70d7b3 AA |
358 | } |
359 | ||
a69aef14 FI |
360 | static int filelayout_write_done_cb(struct rpc_task *task, |
361 | struct nfs_write_data *data) | |
362 | { | |
e7dd79af AA |
363 | struct nfs_pgio_header *hdr = data->header; |
364 | int err; | |
365 | ||
366 | err = filelayout_async_handle_error(task, data->args.context->state, | |
367 | data->ds_clp, hdr->lseg); | |
368 | ||
369 | switch (err) { | |
370 | case -NFS4ERR_RESET_TO_MDS: | |
371 | filelayout_reset_write(data); | |
372 | return task->tk_status; | |
373 | case -EAGAIN: | |
d00c5d43 | 374 | rpc_restart_call_prepare(task); |
a69aef14 FI |
375 | return -EAGAIN; |
376 | } | |
377 | ||
863a3c6c | 378 | filelayout_set_layoutcommit(data); |
a69aef14 FI |
379 | return 0; |
380 | } | |
381 | ||
e0c2b380 | 382 | /* Fake up some data that will cause nfs_commit_release to retry the writes. */ |
0b7c0153 | 383 | static void prepare_to_resend_writes(struct nfs_commit_data *data) |
e0c2b380 FI |
384 | { |
385 | struct nfs_page *first = nfs_list_entry(data->pages.next); | |
386 | ||
387 | data->task.tk_status = 0; | |
2f2c63bc TM |
388 | memcpy(&data->verf.verifier, &first->wb_verf, |
389 | sizeof(data->verf.verifier)); | |
390 | data->verf.verifier.data[0]++; /* ensure verifier mismatch */ | |
e0c2b380 FI |
391 | } |
392 | ||
393 | static int filelayout_commit_done_cb(struct rpc_task *task, | |
0b7c0153 | 394 | struct nfs_commit_data *data) |
e0c2b380 | 395 | { |
e7dd79af AA |
396 | int err; |
397 | ||
398 | err = filelayout_async_handle_error(task, NULL, data->ds_clp, | |
399 | data->lseg); | |
400 | ||
401 | switch (err) { | |
402 | case -NFS4ERR_RESET_TO_MDS: | |
403 | prepare_to_resend_writes(data); | |
404 | return -EAGAIN; | |
405 | case -EAGAIN: | |
406 | rpc_restart_call_prepare(task); | |
e0c2b380 FI |
407 | return -EAGAIN; |
408 | } | |
409 | ||
410 | return 0; | |
411 | } | |
412 | ||
a69aef14 FI |
413 | static void filelayout_write_prepare(struct rpc_task *task, void *data) |
414 | { | |
cd12ae32 | 415 | struct nfs_write_data *wdata = data; |
a69aef14 | 416 | |
c58c8441 TM |
417 | if (unlikely(test_bit(NFS_CONTEXT_BAD, &wdata->args.context->flags))) { |
418 | rpc_exit(task, -EIO); | |
419 | return; | |
420 | } | |
041245c8 | 421 | if (filelayout_reset_to_mds(wdata->header->lseg)) { |
0ad2f378 AA |
422 | dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid); |
423 | filelayout_reset_write(wdata); | |
424 | rpc_exit(task, 0); | |
425 | return; | |
426 | } | |
9b206149 | 427 | if (nfs41_setup_sequence(wdata->ds_clp->cl_session, |
d9afbd1b TM |
428 | &wdata->args.seq_args, |
429 | &wdata->res.seq_res, | |
9b206149 TM |
430 | task)) |
431 | return; | |
432 | nfs4_set_rw_stateid(&wdata->args.stateid, wdata->args.context, | |
433 | wdata->args.lock_context, FMODE_WRITE); | |
a69aef14 FI |
434 | } |
435 | ||
436 | static void filelayout_write_call_done(struct rpc_task *task, void *data) | |
437 | { | |
cd12ae32 | 438 | struct nfs_write_data *wdata = data; |
a69aef14 | 439 | |
bd4aeffb AA |
440 | if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) && |
441 | task->tk_status == 0) | |
0ad2f378 AA |
442 | return; |
443 | ||
a69aef14 | 444 | /* Note this may cause RPC to be resent */ |
cd841605 | 445 | wdata->header->mds_ops->rpc_call_done(task, data); |
a69aef14 FI |
446 | } |
447 | ||
0a702195 WAA |
448 | static void filelayout_write_count_stats(struct rpc_task *task, void *data) |
449 | { | |
cd12ae32 | 450 | struct nfs_write_data *wdata = data; |
0a702195 | 451 | |
cd841605 | 452 | rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics); |
0a702195 WAA |
453 | } |
454 | ||
a69aef14 FI |
455 | static void filelayout_write_release(void *data) |
456 | { | |
cd12ae32 | 457 | struct nfs_write_data *wdata = data; |
d527e5c1 | 458 | struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout; |
a69aef14 | 459 | |
d527e5c1 | 460 | filelayout_fenceme(lo->plh_inode, lo); |
996074cb | 461 | nfs_put_client(wdata->ds_clp); |
cd841605 | 462 | wdata->header->mds_ops->rpc_release(data); |
a69aef14 FI |
463 | } |
464 | ||
0b7c0153 | 465 | static void filelayout_commit_prepare(struct rpc_task *task, void *data) |
e0c2b380 | 466 | { |
0b7c0153 | 467 | struct nfs_commit_data *wdata = data; |
e0c2b380 | 468 | |
d9afbd1b TM |
469 | nfs41_setup_sequence(wdata->ds_clp->cl_session, |
470 | &wdata->args.seq_args, | |
471 | &wdata->res.seq_res, | |
472 | task); | |
0b7c0153 FI |
473 | } |
474 | ||
475 | static void filelayout_write_commit_done(struct rpc_task *task, void *data) | |
476 | { | |
477 | struct nfs_commit_data *wdata = data; | |
478 | ||
479 | /* Note this may cause RPC to be resent */ | |
480 | wdata->mds_ops->rpc_call_done(task, data); | |
481 | } | |
482 | ||
483 | static void filelayout_commit_count_stats(struct rpc_task *task, void *data) | |
484 | { | |
485 | struct nfs_commit_data *cdata = data; | |
486 | ||
487 | rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics); | |
488 | } | |
489 | ||
490 | static void filelayout_commit_release(void *calldata) | |
491 | { | |
492 | struct nfs_commit_data *data = calldata; | |
493 | ||
f453a54a | 494 | data->completion_ops->completion(data); |
9369a431 | 495 | pnfs_put_lseg(data->lseg); |
3a7936c3 | 496 | nfs_put_client(data->ds_clp); |
0b7c0153 | 497 | nfs_commitdata_release(data); |
e0c2b380 FI |
498 | } |
499 | ||
17280175 | 500 | static const struct rpc_call_ops filelayout_read_call_ops = { |
dc70d7b3 AA |
501 | .rpc_call_prepare = filelayout_read_prepare, |
502 | .rpc_call_done = filelayout_read_call_done, | |
0a702195 | 503 | .rpc_count_stats = filelayout_read_count_stats, |
dc70d7b3 AA |
504 | .rpc_release = filelayout_read_release, |
505 | }; | |
506 | ||
17280175 | 507 | static const struct rpc_call_ops filelayout_write_call_ops = { |
a69aef14 FI |
508 | .rpc_call_prepare = filelayout_write_prepare, |
509 | .rpc_call_done = filelayout_write_call_done, | |
0a702195 | 510 | .rpc_count_stats = filelayout_write_count_stats, |
a69aef14 FI |
511 | .rpc_release = filelayout_write_release, |
512 | }; | |
513 | ||
17280175 | 514 | static const struct rpc_call_ops filelayout_commit_call_ops = { |
0b7c0153 FI |
515 | .rpc_call_prepare = filelayout_commit_prepare, |
516 | .rpc_call_done = filelayout_write_commit_done, | |
517 | .rpc_count_stats = filelayout_commit_count_stats, | |
e0c2b380 FI |
518 | .rpc_release = filelayout_commit_release, |
519 | }; | |
520 | ||
dc70d7b3 AA |
521 | static enum pnfs_try_status |
522 | filelayout_read_pagelist(struct nfs_read_data *data) | |
523 | { | |
cd841605 FI |
524 | struct nfs_pgio_header *hdr = data->header; |
525 | struct pnfs_layout_segment *lseg = hdr->lseg; | |
dc70d7b3 AA |
526 | struct nfs4_pnfs_ds *ds; |
527 | loff_t offset = data->args.offset; | |
528 | u32 j, idx; | |
529 | struct nfs_fh *fh; | |
dc70d7b3 AA |
530 | |
531 | dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n", | |
cd841605 | 532 | __func__, hdr->inode->i_ino, |
dc70d7b3 AA |
533 | data->args.pgbase, (size_t)data->args.count, offset); |
534 | ||
535 | /* Retrieve the correct rpc_client for the byte range */ | |
536 | j = nfs4_fl_calc_j_index(lseg, offset); | |
537 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
538 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 539 | if (!ds) |
dc70d7b3 | 540 | return PNFS_NOT_ATTEMPTED; |
3a7936c3 AA |
541 | dprintk("%s USE DS: %s cl_count %d\n", __func__, |
542 | ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count)); | |
dc70d7b3 AA |
543 | |
544 | /* No multipath support. Use first DS */ | |
3a7936c3 | 545 | atomic_inc(&ds->ds_clp->cl_count); |
dc70d7b3 AA |
546 | data->ds_clp = ds->ds_clp; |
547 | fh = nfs4_fl_select_ds_fh(lseg, j); | |
548 | if (fh) | |
549 | data->args.fh = fh; | |
550 | ||
551 | data->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
552 | data->mds_offset = offset; | |
553 | ||
554 | /* Perform an asynchronous read to ds */ | |
bc5a89b3 | 555 | nfs_initiate_read(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 | 556 | &filelayout_read_call_ops, RPC_TASK_SOFTCONN); |
dc70d7b3 AA |
557 | return PNFS_ATTEMPTED; |
558 | } | |
559 | ||
a69aef14 | 560 | /* Perform async writes. */ |
0382b744 AA |
561 | static enum pnfs_try_status |
562 | filelayout_write_pagelist(struct nfs_write_data *data, int sync) | |
563 | { | |
cd841605 FI |
564 | struct nfs_pgio_header *hdr = data->header; |
565 | struct pnfs_layout_segment *lseg = hdr->lseg; | |
a69aef14 FI |
566 | struct nfs4_pnfs_ds *ds; |
567 | loff_t offset = data->args.offset; | |
568 | u32 j, idx; | |
569 | struct nfs_fh *fh; | |
a69aef14 FI |
570 | |
571 | /* Retrieve the correct rpc_client for the byte range */ | |
572 | j = nfs4_fl_calc_j_index(lseg, offset); | |
573 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
574 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 575 | if (!ds) |
a69aef14 | 576 | return PNFS_NOT_ATTEMPTED; |
3a7936c3 AA |
577 | dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n", |
578 | __func__, hdr->inode->i_ino, sync, (size_t) data->args.count, | |
579 | offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count)); | |
a69aef14 | 580 | |
a69aef14 | 581 | data->write_done_cb = filelayout_write_done_cb; |
3a7936c3 | 582 | atomic_inc(&ds->ds_clp->cl_count); |
a69aef14 FI |
583 | data->ds_clp = ds->ds_clp; |
584 | fh = nfs4_fl_select_ds_fh(lseg, j); | |
585 | if (fh) | |
586 | data->args.fh = fh; | |
587 | /* | |
588 | * Get the file offset on the dserver. Set the write offset to | |
589 | * this offset and save the original offset. | |
590 | */ | |
591 | data->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
a69aef14 FI |
592 | |
593 | /* Perform an asynchronous write */ | |
bc5a89b3 | 594 | nfs_initiate_write(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 AA |
595 | &filelayout_write_call_ops, sync, |
596 | RPC_TASK_SOFTCONN); | |
a69aef14 | 597 | return PNFS_ATTEMPTED; |
0382b744 AA |
598 | } |
599 | ||
16b374ca AA |
600 | /* |
601 | * filelayout_check_layout() | |
602 | * | |
603 | * Make sure layout segment parameters are sane WRT the device. | |
604 | * At this point no generic layer initialization of the lseg has occurred, | |
605 | * and nothing has been added to the layout_hdr cache. | |
606 | * | |
607 | */ | |
608 | static int | |
609 | filelayout_check_layout(struct pnfs_layout_hdr *lo, | |
610 | struct nfs4_filelayout_segment *fl, | |
611 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 TM |
612 | struct nfs4_deviceid *id, |
613 | gfp_t gfp_flags) | |
16b374ca | 614 | { |
a1eaecbc | 615 | struct nfs4_deviceid_node *d; |
16b374ca AA |
616 | struct nfs4_file_layout_dsaddr *dsaddr; |
617 | int status = -EINVAL; | |
b7edfaa1 | 618 | struct nfs_server *nfss = NFS_SERVER(lo->plh_inode); |
16b374ca AA |
619 | |
620 | dprintk("--> %s\n", __func__); | |
621 | ||
7c24d948 AA |
622 | /* FIXME: remove this check when layout segment support is added */ |
623 | if (lgr->range.offset != 0 || | |
624 | lgr->range.length != NFS4_MAX_UINT64) { | |
625 | dprintk("%s Only whole file layouts supported. Use MDS i/o\n", | |
626 | __func__); | |
627 | goto out; | |
628 | } | |
629 | ||
16b374ca | 630 | if (fl->pattern_offset > lgr->range.offset) { |
3b6445a6 | 631 | dprintk("%s pattern_offset %lld too large\n", |
16b374ca AA |
632 | __func__, fl->pattern_offset); |
633 | goto out; | |
634 | } | |
635 | ||
75247aff BH |
636 | if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) { |
637 | dprintk("%s Invalid stripe unit (%u)\n", | |
16b374ca AA |
638 | __func__, fl->stripe_unit); |
639 | goto out; | |
640 | } | |
641 | ||
642 | /* find and reference the deviceid */ | |
35c8bb54 BH |
643 | d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld, |
644 | NFS_SERVER(lo->plh_inode)->nfs_client, id); | |
a1eaecbc | 645 | if (d == NULL) { |
cd5875fe TM |
646 | dsaddr = filelayout_get_device_info(lo->plh_inode, id, |
647 | lo->plh_lc_cred, gfp_flags); | |
16b374ca AA |
648 | if (dsaddr == NULL) |
649 | goto out; | |
a1eaecbc BH |
650 | } else |
651 | dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node); | |
1dfed273 TM |
652 | /* Found deviceid is unavailable */ |
653 | if (filelayout_test_devid_unavailable(&dsaddr->id_node)) | |
c47abcf8 AA |
654 | goto out_put; |
655 | ||
16b374ca AA |
656 | fl->dsaddr = dsaddr; |
657 | ||
e414966b CL |
658 | if (fl->first_stripe_index >= dsaddr->stripe_count) { |
659 | dprintk("%s Bad first_stripe_index %u\n", | |
16b374ca AA |
660 | __func__, fl->first_stripe_index); |
661 | goto out_put; | |
662 | } | |
663 | ||
664 | if ((fl->stripe_type == STRIPE_SPARSE && | |
665 | fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) || | |
666 | (fl->stripe_type == STRIPE_DENSE && | |
667 | fl->num_fh != dsaddr->stripe_count)) { | |
668 | dprintk("%s num_fh %u not valid for given packing\n", | |
669 | __func__, fl->num_fh); | |
670 | goto out_put; | |
671 | } | |
672 | ||
673 | if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) { | |
674 | dprintk("%s Stripe unit (%u) not aligned with rsize %u " | |
675 | "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize, | |
676 | nfss->wsize); | |
677 | } | |
678 | ||
679 | status = 0; | |
680 | out: | |
681 | dprintk("--> %s returns %d\n", __func__, status); | |
682 | return status; | |
683 | out_put: | |
ea8eecdd | 684 | nfs4_fl_put_deviceid(dsaddr); |
16b374ca AA |
685 | goto out; |
686 | } | |
687 | ||
688 | static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl) | |
689 | { | |
690 | int i; | |
691 | ||
692 | for (i = 0; i < fl->num_fh; i++) { | |
693 | if (!fl->fh_array[i]) | |
694 | break; | |
695 | kfree(fl->fh_array[i]); | |
696 | } | |
697 | kfree(fl->fh_array); | |
698 | fl->fh_array = NULL; | |
699 | } | |
700 | ||
701 | static void | |
702 | _filelayout_free_lseg(struct nfs4_filelayout_segment *fl) | |
703 | { | |
704 | filelayout_free_fh_array(fl); | |
705 | kfree(fl); | |
706 | } | |
707 | ||
708 | static int | |
709 | filelayout_decode_layout(struct pnfs_layout_hdr *flo, | |
710 | struct nfs4_filelayout_segment *fl, | |
711 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 TM |
712 | struct nfs4_deviceid *id, |
713 | gfp_t gfp_flags) | |
16b374ca | 714 | { |
35124a09 | 715 | struct xdr_stream stream; |
f7da7a12 | 716 | struct xdr_buf buf; |
35124a09 WAA |
717 | struct page *scratch; |
718 | __be32 *p; | |
16b374ca AA |
719 | uint32_t nfl_util; |
720 | int i; | |
721 | ||
722 | dprintk("%s: set_layout_map Begin\n", __func__); | |
723 | ||
a75b9df9 | 724 | scratch = alloc_page(gfp_flags); |
35124a09 WAA |
725 | if (!scratch) |
726 | return -ENOMEM; | |
727 | ||
f7da7a12 | 728 | xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len); |
35124a09 WAA |
729 | xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE); |
730 | ||
731 | /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8), | |
732 | * num_fh (4) */ | |
733 | p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20); | |
734 | if (unlikely(!p)) | |
735 | goto out_err; | |
736 | ||
16b374ca AA |
737 | memcpy(id, p, sizeof(*id)); |
738 | p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE); | |
a1eaecbc | 739 | nfs4_print_deviceid(id); |
16b374ca AA |
740 | |
741 | nfl_util = be32_to_cpup(p++); | |
742 | if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS) | |
743 | fl->commit_through_mds = 1; | |
744 | if (nfl_util & NFL4_UFLG_DENSE) | |
745 | fl->stripe_type = STRIPE_DENSE; | |
746 | else | |
747 | fl->stripe_type = STRIPE_SPARSE; | |
748 | fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK; | |
749 | ||
750 | fl->first_stripe_index = be32_to_cpup(p++); | |
751 | p = xdr_decode_hyper(p, &fl->pattern_offset); | |
752 | fl->num_fh = be32_to_cpup(p++); | |
753 | ||
754 | dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n", | |
755 | __func__, nfl_util, fl->num_fh, fl->first_stripe_index, | |
756 | fl->pattern_offset); | |
757 | ||
cec765cf AA |
758 | /* Note that a zero value for num_fh is legal for STRIPE_SPARSE. |
759 | * Futher checking is done in filelayout_check_layout */ | |
e414966b | 760 | if (fl->num_fh > |
cec765cf | 761 | max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT)) |
35124a09 WAA |
762 | goto out_err; |
763 | ||
cec765cf | 764 | if (fl->num_fh > 0) { |
1813badd | 765 | fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]), |
cec765cf AA |
766 | gfp_flags); |
767 | if (!fl->fh_array) | |
768 | goto out_err; | |
769 | } | |
16b374ca AA |
770 | |
771 | for (i = 0; i < fl->num_fh; i++) { | |
772 | /* Do we want to use a mempool here? */ | |
a75b9df9 | 773 | fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags); |
35124a09 WAA |
774 | if (!fl->fh_array[i]) |
775 | goto out_err_free; | |
776 | ||
777 | p = xdr_inline_decode(&stream, 4); | |
778 | if (unlikely(!p)) | |
779 | goto out_err_free; | |
16b374ca AA |
780 | fl->fh_array[i]->size = be32_to_cpup(p++); |
781 | if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) { | |
f9fd2d9c | 782 | printk(KERN_ERR "NFS: Too big fh %d received %d\n", |
16b374ca | 783 | i, fl->fh_array[i]->size); |
35124a09 | 784 | goto out_err_free; |
16b374ca | 785 | } |
35124a09 WAA |
786 | |
787 | p = xdr_inline_decode(&stream, fl->fh_array[i]->size); | |
788 | if (unlikely(!p)) | |
789 | goto out_err_free; | |
16b374ca | 790 | memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size); |
16b374ca AA |
791 | dprintk("DEBUG: %s: fh len %d\n", __func__, |
792 | fl->fh_array[i]->size); | |
793 | } | |
794 | ||
35124a09 | 795 | __free_page(scratch); |
16b374ca | 796 | return 0; |
35124a09 WAA |
797 | |
798 | out_err_free: | |
799 | filelayout_free_fh_array(fl); | |
800 | out_err: | |
801 | __free_page(scratch); | |
802 | return -EIO; | |
16b374ca AA |
803 | } |
804 | ||
c879513e FI |
805 | static void |
806 | filelayout_free_lseg(struct pnfs_layout_segment *lseg) | |
807 | { | |
808 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
809 | ||
810 | dprintk("--> %s\n", __func__); | |
811 | nfs4_fl_put_deviceid(fl->dsaddr); | |
799ba8d5 FI |
812 | /* This assumes a single RW lseg */ |
813 | if (lseg->pls_range.iomode == IOMODE_RW) { | |
814 | struct nfs4_filelayout *flo; | |
815 | ||
816 | flo = FILELAYOUT_FROM_HDR(lseg->pls_layout); | |
817 | flo->commit_info.nbuckets = 0; | |
818 | kfree(flo->commit_info.buckets); | |
819 | flo->commit_info.buckets = NULL; | |
820 | } | |
c879513e FI |
821 | _filelayout_free_lseg(fl); |
822 | } | |
823 | ||
799ba8d5 FI |
824 | static int |
825 | filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg, | |
ea2cf228 | 826 | struct nfs_commit_info *cinfo, |
799ba8d5 FI |
827 | gfp_t gfp_flags) |
828 | { | |
829 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
ea2cf228 | 830 | struct pnfs_commit_bucket *buckets; |
799ba8d5 FI |
831 | int size; |
832 | ||
833 | if (fl->commit_through_mds) | |
834 | return 0; | |
ea2cf228 | 835 | if (cinfo->ds->nbuckets != 0) { |
799ba8d5 FI |
836 | /* This assumes there is only one IOMODE_RW lseg. What |
837 | * we really want to do is have a layout_hdr level | |
838 | * dictionary of <multipath_list4, fh> keys, each | |
839 | * associated with a struct list_head, populated by calls | |
840 | * to filelayout_write_pagelist(). | |
841 | * */ | |
842 | return 0; | |
843 | } | |
844 | ||
845 | size = (fl->stripe_type == STRIPE_SPARSE) ? | |
846 | fl->dsaddr->ds_num : fl->dsaddr->stripe_count; | |
847 | ||
ea2cf228 | 848 | buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket), |
799ba8d5 FI |
849 | gfp_flags); |
850 | if (!buckets) | |
851 | return -ENOMEM; | |
852 | else { | |
853 | int i; | |
854 | ||
ea2cf228 FI |
855 | spin_lock(cinfo->lock); |
856 | if (cinfo->ds->nbuckets != 0) | |
799ba8d5 FI |
857 | kfree(buckets); |
858 | else { | |
ea2cf228 FI |
859 | cinfo->ds->buckets = buckets; |
860 | cinfo->ds->nbuckets = size; | |
799ba8d5 FI |
861 | for (i = 0; i < size; i++) { |
862 | INIT_LIST_HEAD(&buckets[i].written); | |
863 | INIT_LIST_HEAD(&buckets[i].committing); | |
864 | } | |
865 | } | |
ea2cf228 | 866 | spin_unlock(cinfo->lock); |
799ba8d5 FI |
867 | return 0; |
868 | } | |
869 | } | |
870 | ||
16b374ca AA |
871 | static struct pnfs_layout_segment * |
872 | filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid, | |
a75b9df9 TM |
873 | struct nfs4_layoutget_res *lgr, |
874 | gfp_t gfp_flags) | |
16b374ca AA |
875 | { |
876 | struct nfs4_filelayout_segment *fl; | |
877 | int rc; | |
878 | struct nfs4_deviceid id; | |
879 | ||
880 | dprintk("--> %s\n", __func__); | |
a75b9df9 | 881 | fl = kzalloc(sizeof(*fl), gfp_flags); |
16b374ca AA |
882 | if (!fl) |
883 | return NULL; | |
884 | ||
a75b9df9 TM |
885 | rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags); |
886 | if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) { | |
16b374ca AA |
887 | _filelayout_free_lseg(fl); |
888 | return NULL; | |
889 | } | |
890 | return &fl->generic_hdr; | |
891 | } | |
892 | ||
94ad1c80 FI |
893 | /* |
894 | * filelayout_pg_test(). Called by nfs_can_coalesce_requests() | |
895 | * | |
18ad0a9f BH |
896 | * return true : coalesce page |
897 | * return false : don't coalesce page | |
94ad1c80 | 898 | */ |
1751c363 | 899 | static bool |
94ad1c80 FI |
900 | filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev, |
901 | struct nfs_page *req) | |
902 | { | |
903 | u64 p_stripe, r_stripe; | |
904 | u32 stripe_unit; | |
905 | ||
19345cb2 BH |
906 | if (!pnfs_generic_pg_test(pgio, prev, req) || |
907 | !nfs_generic_pg_test(pgio, prev, req)) | |
908 | return false; | |
89a58e32 | 909 | |
b5542849 FI |
910 | p_stripe = (u64)req_offset(prev); |
911 | r_stripe = (u64)req_offset(req); | |
94ad1c80 FI |
912 | stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit; |
913 | ||
914 | do_div(p_stripe, stripe_unit); | |
915 | do_div(r_stripe, stripe_unit); | |
916 | ||
917 | return (p_stripe == r_stripe); | |
918 | } | |
919 | ||
17280175 | 920 | static void |
7c24d948 AA |
921 | filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio, |
922 | struct nfs_page *req) | |
923 | { | |
bc5a89b3 | 924 | WARN_ON_ONCE(pgio->pg_lseg != NULL); |
7c24d948 | 925 | |
1825a0d0 FI |
926 | if (req->wb_offset != req->wb_pgbase) { |
927 | /* | |
928 | * Handling unaligned pages is difficult, because have to | |
929 | * somehow split a req in two in certain cases in the | |
930 | * pg.test code. Avoid this by just not using pnfs | |
931 | * in this case. | |
932 | */ | |
933 | nfs_pageio_reset_read_mds(pgio); | |
934 | return; | |
935 | } | |
7c24d948 AA |
936 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
937 | req->wb_context, | |
938 | 0, | |
939 | NFS4_MAX_UINT64, | |
940 | IOMODE_READ, | |
941 | GFP_KERNEL); | |
942 | /* If no lseg, fall back to read through mds */ | |
943 | if (pgio->pg_lseg == NULL) | |
1f945357 | 944 | nfs_pageio_reset_read_mds(pgio); |
7c24d948 AA |
945 | } |
946 | ||
17280175 | 947 | static void |
7c24d948 AA |
948 | filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio, |
949 | struct nfs_page *req) | |
950 | { | |
ea2cf228 | 951 | struct nfs_commit_info cinfo; |
799ba8d5 FI |
952 | int status; |
953 | ||
bc5a89b3 | 954 | WARN_ON_ONCE(pgio->pg_lseg != NULL); |
7c24d948 | 955 | |
1825a0d0 FI |
956 | if (req->wb_offset != req->wb_pgbase) |
957 | goto out_mds; | |
7c24d948 AA |
958 | pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode, |
959 | req->wb_context, | |
960 | 0, | |
961 | NFS4_MAX_UINT64, | |
962 | IOMODE_RW, | |
963 | GFP_NOFS); | |
964 | /* If no lseg, fall back to write through mds */ | |
965 | if (pgio->pg_lseg == NULL) | |
799ba8d5 | 966 | goto out_mds; |
ea2cf228 FI |
967 | nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq); |
968 | status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS); | |
799ba8d5 | 969 | if (status < 0) { |
9369a431 | 970 | pnfs_put_lseg(pgio->pg_lseg); |
799ba8d5 FI |
971 | pgio->pg_lseg = NULL; |
972 | goto out_mds; | |
973 | } | |
974 | return; | |
975 | out_mds: | |
976 | nfs_pageio_reset_write_mds(pgio); | |
7c24d948 AA |
977 | } |
978 | ||
1751c363 | 979 | static const struct nfs_pageio_ops filelayout_pg_read_ops = { |
7c24d948 | 980 | .pg_init = filelayout_pg_init_read, |
1751c363 | 981 | .pg_test = filelayout_pg_test, |
493292dd | 982 | .pg_doio = pnfs_generic_pg_readpages, |
1751c363 TM |
983 | }; |
984 | ||
985 | static const struct nfs_pageio_ops filelayout_pg_write_ops = { | |
7c24d948 | 986 | .pg_init = filelayout_pg_init_write, |
1751c363 | 987 | .pg_test = filelayout_pg_test, |
dce81290 | 988 | .pg_doio = pnfs_generic_pg_writepages, |
1751c363 TM |
989 | }; |
990 | ||
e0c2b380 FI |
991 | static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j) |
992 | { | |
993 | if (fl->stripe_type == STRIPE_SPARSE) | |
994 | return nfs4_fl_calc_ds_index(&fl->generic_hdr, j); | |
995 | else | |
996 | return j; | |
997 | } | |
998 | ||
d6d6dc7c FI |
999 | /* The generic layer is about to remove the req from the commit list. |
1000 | * If this will make the bucket empty, it will need to put the lseg reference. | |
d6d6dc7c | 1001 | */ |
8dd37758 | 1002 | static void |
ea2cf228 FI |
1003 | filelayout_clear_request_commit(struct nfs_page *req, |
1004 | struct nfs_commit_info *cinfo) | |
d6d6dc7c | 1005 | { |
8dd37758 | 1006 | struct pnfs_layout_segment *freeme = NULL; |
8dd37758 | 1007 | |
ea2cf228 | 1008 | spin_lock(cinfo->lock); |
8dd37758 TM |
1009 | if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags)) |
1010 | goto out; | |
ea2cf228 | 1011 | cinfo->ds->nwritten--; |
d6d6dc7c | 1012 | if (list_is_singular(&req->wb_list)) { |
ea2cf228 | 1013 | struct pnfs_commit_bucket *bucket; |
d6d6dc7c | 1014 | |
799ba8d5 | 1015 | bucket = list_first_entry(&req->wb_list, |
ea2cf228 | 1016 | struct pnfs_commit_bucket, |
799ba8d5 FI |
1017 | written); |
1018 | freeme = bucket->wlseg; | |
1019 | bucket->wlseg = NULL; | |
d6d6dc7c | 1020 | } |
8dd37758 | 1021 | out: |
ea2cf228 FI |
1022 | nfs_request_remove_commit_list(req, cinfo); |
1023 | spin_unlock(cinfo->lock); | |
9369a431 | 1024 | pnfs_put_lseg(freeme); |
d6d6dc7c FI |
1025 | } |
1026 | ||
1027 | static struct list_head * | |
1028 | filelayout_choose_commit_list(struct nfs_page *req, | |
ea2cf228 FI |
1029 | struct pnfs_layout_segment *lseg, |
1030 | struct nfs_commit_info *cinfo) | |
e0c2b380 | 1031 | { |
e0c2b380 FI |
1032 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); |
1033 | u32 i, j; | |
1034 | struct list_head *list; | |
ea2cf228 | 1035 | struct pnfs_commit_bucket *buckets; |
e0c2b380 | 1036 | |
d6d6dc7c | 1037 | if (fl->commit_through_mds) |
ea2cf228 | 1038 | return &cinfo->mds->list; |
d6d6dc7c | 1039 | |
e0c2b380 FI |
1040 | /* Note that we are calling nfs4_fl_calc_j_index on each page |
1041 | * that ends up being committed to a data server. An attractive | |
1042 | * alternative is to add a field to nfs_write_data and nfs_page | |
1043 | * to store the value calculated in filelayout_write_pagelist | |
1044 | * and just use that here. | |
1045 | */ | |
b5542849 | 1046 | j = nfs4_fl_calc_j_index(lseg, req_offset(req)); |
e0c2b380 | 1047 | i = select_bucket_index(fl, j); |
ea2cf228 | 1048 | buckets = cinfo->ds->buckets; |
799ba8d5 | 1049 | list = &buckets[i].written; |
e0c2b380 | 1050 | if (list_empty(list)) { |
d6d6dc7c FI |
1051 | /* Non-empty buckets hold a reference on the lseg. That ref |
1052 | * is normally transferred to the COMMIT call and released | |
1053 | * there. It could also be released if the last req is pulled | |
1054 | * off due to a rewrite, in which case it will be done in | |
799ba8d5 | 1055 | * filelayout_clear_request_commit |
d6d6dc7c | 1056 | */ |
9369a431 | 1057 | buckets[i].wlseg = pnfs_get_lseg(lseg); |
e0c2b380 | 1058 | } |
8dd37758 | 1059 | set_bit(PG_COMMIT_TO_DS, &req->wb_flags); |
ea2cf228 | 1060 | cinfo->ds->nwritten++; |
e0c2b380 FI |
1061 | return list; |
1062 | } | |
1063 | ||
8dd37758 TM |
1064 | static void |
1065 | filelayout_mark_request_commit(struct nfs_page *req, | |
ea2cf228 FI |
1066 | struct pnfs_layout_segment *lseg, |
1067 | struct nfs_commit_info *cinfo) | |
8dd37758 TM |
1068 | { |
1069 | struct list_head *list; | |
1070 | ||
ea2cf228 FI |
1071 | list = filelayout_choose_commit_list(req, lseg, cinfo); |
1072 | nfs_request_add_commit_list(req, list, cinfo); | |
8dd37758 TM |
1073 | } |
1074 | ||
e0c2b380 FI |
1075 | static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i) |
1076 | { | |
1077 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1078 | ||
1079 | if (flseg->stripe_type == STRIPE_SPARSE) | |
1080 | return i; | |
1081 | else | |
1082 | return nfs4_fl_calc_ds_index(lseg, i); | |
1083 | } | |
1084 | ||
1085 | static struct nfs_fh * | |
1086 | select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i) | |
1087 | { | |
1088 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1089 | ||
1090 | if (flseg->stripe_type == STRIPE_SPARSE) { | |
1091 | if (flseg->num_fh == 1) | |
1092 | i = 0; | |
1093 | else if (flseg->num_fh == 0) | |
1094 | /* Use the MDS OPEN fh set in nfs_read_rpcsetup */ | |
1095 | return NULL; | |
1096 | } | |
1097 | return flseg->fh_array[i]; | |
1098 | } | |
1099 | ||
0b7c0153 | 1100 | static int filelayout_initiate_commit(struct nfs_commit_data *data, int how) |
e0c2b380 FI |
1101 | { |
1102 | struct pnfs_layout_segment *lseg = data->lseg; | |
1103 | struct nfs4_pnfs_ds *ds; | |
1104 | u32 idx; | |
1105 | struct nfs_fh *fh; | |
1106 | ||
1107 | idx = calc_ds_index_from_commit(lseg, data->ds_commit_index); | |
1108 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
1109 | if (!ds) { | |
e0c2b380 | 1110 | prepare_to_resend_writes(data); |
8dd37758 | 1111 | filelayout_commit_release(data); |
e0c2b380 FI |
1112 | return -EAGAIN; |
1113 | } | |
3a7936c3 AA |
1114 | dprintk("%s ino %lu, how %d cl_count %d\n", __func__, |
1115 | data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count)); | |
0b7c0153 | 1116 | data->commit_done_cb = filelayout_commit_done_cb; |
3a7936c3 | 1117 | atomic_inc(&ds->ds_clp->cl_count); |
e0c2b380 FI |
1118 | data->ds_clp = ds->ds_clp; |
1119 | fh = select_ds_fh_from_commit(lseg, data->ds_commit_index); | |
1120 | if (fh) | |
1121 | data->args.fh = fh; | |
0b7c0153 | 1122 | return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data, |
9f0ec176 AA |
1123 | &filelayout_commit_call_ops, how, |
1124 | RPC_TASK_SOFTCONN); | |
e0c2b380 FI |
1125 | } |
1126 | ||
8dd37758 | 1127 | static int |
1763da12 FI |
1128 | transfer_commit_list(struct list_head *src, struct list_head *dst, |
1129 | struct nfs_commit_info *cinfo, int max) | |
8dd37758 | 1130 | { |
8dd37758 TM |
1131 | struct nfs_page *req, *tmp; |
1132 | int ret = 0; | |
1133 | ||
1134 | list_for_each_entry_safe(req, tmp, src, wb_list) { | |
1135 | if (!nfs_lock_request(req)) | |
1136 | continue; | |
53b8ee34 | 1137 | kref_get(&req->wb_kref); |
ea2cf228 | 1138 | if (cond_resched_lock(cinfo->lock)) |
3b3be88d | 1139 | list_safe_reset_next(req, tmp, wb_list); |
ea2cf228 | 1140 | nfs_request_remove_commit_list(req, cinfo); |
8dd37758 TM |
1141 | clear_bit(PG_COMMIT_TO_DS, &req->wb_flags); |
1142 | nfs_list_add_request(req, dst); | |
1143 | ret++; | |
1763da12 | 1144 | if ((ret == max) && !cinfo->dreq) |
8dd37758 TM |
1145 | break; |
1146 | } | |
1763da12 FI |
1147 | return ret; |
1148 | } | |
1149 | ||
1150 | static int | |
1151 | filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket, | |
1152 | struct nfs_commit_info *cinfo, | |
1153 | int max) | |
1154 | { | |
1155 | struct list_head *src = &bucket->written; | |
1156 | struct list_head *dst = &bucket->committing; | |
1157 | int ret; | |
1158 | ||
1159 | ret = transfer_commit_list(src, dst, cinfo, max); | |
799ba8d5 | 1160 | if (ret) { |
ea2cf228 FI |
1161 | cinfo->ds->nwritten -= ret; |
1162 | cinfo->ds->ncommitting += ret; | |
799ba8d5 FI |
1163 | bucket->clseg = bucket->wlseg; |
1164 | if (list_empty(src)) | |
1165 | bucket->wlseg = NULL; | |
1166 | else | |
9369a431 | 1167 | pnfs_get_lseg(bucket->clseg); |
799ba8d5 | 1168 | } |
8dd37758 TM |
1169 | return ret; |
1170 | } | |
1171 | ||
d6d6dc7c | 1172 | /* Move reqs from written to committing lists, returning count of number moved. |
ea2cf228 | 1173 | * Note called with cinfo->lock held. |
d6d6dc7c | 1174 | */ |
ea2cf228 FI |
1175 | static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo, |
1176 | int max) | |
d6d6dc7c | 1177 | { |
d6d6dc7c FI |
1178 | int i, rv = 0, cnt; |
1179 | ||
ea2cf228 FI |
1180 | for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) { |
1181 | cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i], | |
1182 | cinfo, max); | |
d6d6dc7c FI |
1183 | max -= cnt; |
1184 | rv += cnt; | |
1185 | } | |
d6d6dc7c FI |
1186 | return rv; |
1187 | } | |
1188 | ||
1763da12 FI |
1189 | /* Pull everything off the committing lists and dump into @dst */ |
1190 | static void filelayout_recover_commit_reqs(struct list_head *dst, | |
1191 | struct nfs_commit_info *cinfo) | |
1192 | { | |
1193 | struct pnfs_commit_bucket *b; | |
1194 | int i; | |
1195 | ||
1196 | /* NOTE cinfo->lock is NOT held, relying on fact that this is | |
1197 | * only called on single thread per dreq. | |
9369a431 | 1198 | * Can't take the lock because need to do pnfs_put_lseg |
1763da12 FI |
1199 | */ |
1200 | for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) { | |
1201 | if (transfer_commit_list(&b->written, dst, cinfo, 0)) { | |
9369a431 | 1202 | pnfs_put_lseg(b->wlseg); |
1763da12 FI |
1203 | b->wlseg = NULL; |
1204 | } | |
1205 | } | |
1206 | cinfo->ds->nwritten = 0; | |
1207 | } | |
1208 | ||
9390f425 | 1209 | static unsigned int |
ea2cf228 | 1210 | alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list) |
e0c2b380 | 1211 | { |
ea2cf228 FI |
1212 | struct pnfs_ds_commit_info *fl_cinfo; |
1213 | struct pnfs_commit_bucket *bucket; | |
0b7c0153 | 1214 | struct nfs_commit_data *data; |
e0c2b380 | 1215 | int i, j; |
9390f425 | 1216 | unsigned int nreq = 0; |
e0c2b380 | 1217 | |
ea2cf228 | 1218 | fl_cinfo = cinfo->ds; |
799ba8d5 FI |
1219 | bucket = fl_cinfo->buckets; |
1220 | for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) { | |
1221 | if (list_empty(&bucket->committing)) | |
e0c2b380 FI |
1222 | continue; |
1223 | data = nfs_commitdata_alloc(); | |
1224 | if (!data) | |
9390f425 | 1225 | break; |
e0c2b380 | 1226 | data->ds_commit_index = i; |
799ba8d5 FI |
1227 | data->lseg = bucket->clseg; |
1228 | bucket->clseg = NULL; | |
e0c2b380 | 1229 | list_add(&data->pages, list); |
9390f425 | 1230 | nreq++; |
e0c2b380 | 1231 | } |
e0c2b380 | 1232 | |
9390f425 | 1233 | /* Clean up on error */ |
799ba8d5 FI |
1234 | for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) { |
1235 | if (list_empty(&bucket->committing)) | |
e0c2b380 | 1236 | continue; |
ea2cf228 | 1237 | nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo); |
9369a431 | 1238 | pnfs_put_lseg(bucket->clseg); |
799ba8d5 | 1239 | bucket->clseg = NULL; |
e0c2b380 | 1240 | } |
e0c2b380 | 1241 | /* Caller will clean up entries put on list */ |
9390f425 | 1242 | return nreq; |
e0c2b380 FI |
1243 | } |
1244 | ||
1245 | /* This follows nfs_commit_list pretty closely */ | |
1246 | static int | |
1247 | filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages, | |
ea2cf228 | 1248 | int how, struct nfs_commit_info *cinfo) |
e0c2b380 | 1249 | { |
0b7c0153 | 1250 | struct nfs_commit_data *data, *tmp; |
e0c2b380 | 1251 | LIST_HEAD(list); |
9390f425 | 1252 | unsigned int nreq = 0; |
e0c2b380 FI |
1253 | |
1254 | if (!list_empty(mds_pages)) { | |
1255 | data = nfs_commitdata_alloc(); | |
9390f425 TM |
1256 | if (data != NULL) { |
1257 | data->lseg = NULL; | |
1258 | list_add(&data->pages, &list); | |
1259 | nreq++; | |
1260 | } else | |
ea2cf228 | 1261 | nfs_retry_commit(mds_pages, NULL, cinfo); |
e0c2b380 FI |
1262 | } |
1263 | ||
ea2cf228 | 1264 | nreq += alloc_ds_commits(cinfo, &list); |
9390f425 TM |
1265 | |
1266 | if (nreq == 0) { | |
f453a54a | 1267 | cinfo->completion_ops->error_cleanup(NFS_I(inode)); |
9390f425 TM |
1268 | goto out; |
1269 | } | |
1270 | ||
ea2cf228 | 1271 | atomic_add(nreq, &cinfo->mds->rpcs_out); |
e0c2b380 FI |
1272 | |
1273 | list_for_each_entry_safe(data, tmp, &list, pages) { | |
1274 | list_del_init(&data->pages); | |
e0c2b380 | 1275 | if (!data->lseg) { |
f453a54a | 1276 | nfs_init_commit(data, mds_pages, NULL, cinfo); |
0b7c0153 | 1277 | nfs_initiate_commit(NFS_CLIENT(inode), data, |
9f0ec176 | 1278 | data->mds_ops, how, 0); |
e0c2b380 | 1279 | } else { |
ea2cf228 | 1280 | struct pnfs_commit_bucket *buckets; |
799ba8d5 | 1281 | |
ea2cf228 | 1282 | buckets = cinfo->ds->buckets; |
f453a54a | 1283 | nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo); |
e0c2b380 FI |
1284 | filelayout_initiate_commit(data, how); |
1285 | } | |
1286 | } | |
9390f425 | 1287 | out: |
ea2cf228 | 1288 | cinfo->ds->ncommitting = 0; |
9390f425 | 1289 | return PNFS_ATTEMPTED; |
e0c2b380 FI |
1290 | } |
1291 | ||
1775bc34 BH |
1292 | static void |
1293 | filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d) | |
1294 | { | |
1295 | nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node)); | |
1296 | } | |
1297 | ||
799ba8d5 FI |
1298 | static struct pnfs_layout_hdr * |
1299 | filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags) | |
1300 | { | |
1301 | struct nfs4_filelayout *flo; | |
1302 | ||
1303 | flo = kzalloc(sizeof(*flo), gfp_flags); | |
1304 | return &flo->generic_hdr; | |
1305 | } | |
1306 | ||
1307 | static void | |
1308 | filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo) | |
1309 | { | |
1310 | kfree(FILELAYOUT_FROM_HDR(lo)); | |
1311 | } | |
1312 | ||
ea2cf228 FI |
1313 | static struct pnfs_ds_commit_info * |
1314 | filelayout_get_ds_info(struct inode *inode) | |
1315 | { | |
df011748 FI |
1316 | struct pnfs_layout_hdr *layout = NFS_I(inode)->layout; |
1317 | ||
1318 | if (layout == NULL) | |
1319 | return NULL; | |
1320 | else | |
1321 | return &FILELAYOUT_FROM_HDR(layout)->commit_info; | |
ea2cf228 FI |
1322 | } |
1323 | ||
7ab672ce | 1324 | static struct pnfs_layoutdriver_type filelayout_type = { |
ea8eecdd CH |
1325 | .id = LAYOUT_NFSV4_1_FILES, |
1326 | .name = "LAYOUT_NFSV4_1_FILES", | |
1327 | .owner = THIS_MODULE, | |
799ba8d5 FI |
1328 | .alloc_layout_hdr = filelayout_alloc_layout_hdr, |
1329 | .free_layout_hdr = filelayout_free_layout_hdr, | |
ea8eecdd CH |
1330 | .alloc_lseg = filelayout_alloc_lseg, |
1331 | .free_lseg = filelayout_free_lseg, | |
1751c363 TM |
1332 | .pg_read_ops = &filelayout_pg_read_ops, |
1333 | .pg_write_ops = &filelayout_pg_write_ops, | |
ea2cf228 | 1334 | .get_ds_info = &filelayout_get_ds_info, |
8dd37758 TM |
1335 | .mark_request_commit = filelayout_mark_request_commit, |
1336 | .clear_request_commit = filelayout_clear_request_commit, | |
d6d6dc7c | 1337 | .scan_commit_lists = filelayout_scan_commit_lists, |
1763da12 | 1338 | .recover_commit_reqs = filelayout_recover_commit_reqs, |
e0c2b380 | 1339 | .commit_pagelist = filelayout_commit_pagelist, |
dc70d7b3 | 1340 | .read_pagelist = filelayout_read_pagelist, |
0382b744 | 1341 | .write_pagelist = filelayout_write_pagelist, |
1775bc34 | 1342 | .free_deviceid_node = filelayout_free_deveiceid_node, |
7ab672ce DH |
1343 | }; |
1344 | ||
1345 | static int __init nfs4filelayout_init(void) | |
1346 | { | |
1347 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n", | |
1348 | __func__); | |
1349 | return pnfs_register_layoutdriver(&filelayout_type); | |
1350 | } | |
1351 | ||
1352 | static void __exit nfs4filelayout_exit(void) | |
1353 | { | |
1354 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n", | |
1355 | __func__); | |
1356 | pnfs_unregister_layoutdriver(&filelayout_type); | |
1357 | } | |
1358 | ||
f85ef69c BF |
1359 | MODULE_ALIAS("nfs-layouttype4-1"); |
1360 | ||
7ab672ce DH |
1361 | module_init(nfs4filelayout_init); |
1362 | module_exit(nfs4filelayout_exit); |