]>
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> |
66114cad | 35 | #include <linux/backing-dev.h> |
16b374ca | 36 | |
0a702195 WAA |
37 | #include <linux/sunrpc/metrics.h> |
38 | ||
b5968725 TH |
39 | #include "../nfs4session.h" |
40 | #include "../internal.h" | |
41 | #include "../delegation.h" | |
42 | #include "filelayout.h" | |
43 | #include "../nfs4trace.h" | |
7ab672ce DH |
44 | |
45 | #define NFSDBG_FACILITY NFSDBG_PNFS_LD | |
46 | ||
47 | MODULE_LICENSE("GPL"); | |
48 | MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>"); | |
49 | MODULE_DESCRIPTION("The NFSv4 file layout driver"); | |
50 | ||
cbdabc7f AA |
51 | #define FILELAYOUT_POLL_RETRY_MAX (15*HZ) |
52 | ||
cfe7f412 FI |
53 | static loff_t |
54 | filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg, | |
55 | loff_t offset) | |
56 | { | |
57 | u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count; | |
3476f114 CM |
58 | u64 stripe_no; |
59 | u32 rem; | |
cfe7f412 FI |
60 | |
61 | offset -= flseg->pattern_offset; | |
3476f114 CM |
62 | stripe_no = div_u64(offset, stripe_width); |
63 | div_u64_rem(offset, flseg->stripe_unit, &rem); | |
cfe7f412 | 64 | |
3476f114 | 65 | return stripe_no * flseg->stripe_unit + rem; |
cfe7f412 FI |
66 | } |
67 | ||
68 | /* This function is used by the layout driver to calculate the | |
69 | * offset of the file on the dserver based on whether the | |
70 | * layout type is STRIPE_DENSE or STRIPE_SPARSE | |
71 | */ | |
72 | static loff_t | |
73 | filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset) | |
74 | { | |
75 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
76 | ||
77 | switch (flseg->stripe_type) { | |
78 | case STRIPE_SPARSE: | |
79 | return offset; | |
80 | ||
81 | case STRIPE_DENSE: | |
82 | return filelayout_get_dense_offset(flseg, offset); | |
83 | } | |
84 | ||
85 | BUG(); | |
86 | } | |
87 | ||
d45f60c6 | 88 | static void filelayout_reset_write(struct nfs_pgio_header *hdr) |
e7dd79af | 89 | { |
d45f60c6 | 90 | struct rpc_task *task = &hdr->task; |
e7dd79af AA |
91 | |
92 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { | |
93 | dprintk("%s Reset task %5u for i/o through MDS " | |
1e8968c5 | 94 | "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, |
d45f60c6 | 95 | hdr->task.tk_pid, |
497826af | 96 | hdr->inode->i_sb->s_id, |
1e8968c5 | 97 | (unsigned long long)NFS_FILEID(hdr->inode), |
d45f60c6 WAA |
98 | hdr->args.count, |
99 | (unsigned long long)hdr->args.offset); | |
e7dd79af | 100 | |
53113ad3 | 101 | task->tk_status = pnfs_write_done_resend_to_mds(hdr); |
e7dd79af AA |
102 | } |
103 | } | |
104 | ||
d45f60c6 | 105 | static void filelayout_reset_read(struct nfs_pgio_header *hdr) |
e7dd79af | 106 | { |
d45f60c6 | 107 | struct rpc_task *task = &hdr->task; |
e7dd79af AA |
108 | |
109 | if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) { | |
110 | dprintk("%s Reset task %5u for i/o through MDS " | |
1e8968c5 | 111 | "(req %s/%llu, %u bytes @ offset %llu)\n", __func__, |
d45f60c6 | 112 | hdr->task.tk_pid, |
497826af | 113 | hdr->inode->i_sb->s_id, |
1e8968c5 | 114 | (unsigned long long)NFS_FILEID(hdr->inode), |
d45f60c6 WAA |
115 | hdr->args.count, |
116 | (unsigned long long)hdr->args.offset); | |
e7dd79af | 117 | |
53113ad3 | 118 | task->tk_status = pnfs_read_done_resend_to_mds(hdr); |
e7dd79af AA |
119 | } |
120 | } | |
121 | ||
cbdabc7f AA |
122 | static int filelayout_async_handle_error(struct rpc_task *task, |
123 | struct nfs4_state *state, | |
124 | struct nfs_client *clp, | |
e7dd79af | 125 | struct pnfs_layout_segment *lseg) |
cbdabc7f | 126 | { |
d527e5c1 TM |
127 | struct pnfs_layout_hdr *lo = lseg->pls_layout; |
128 | struct inode *inode = lo->plh_inode; | |
e7dd79af | 129 | struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg); |
671fb896 | 130 | struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table; |
9cb81968 | 131 | |
cbdabc7f AA |
132 | if (task->tk_status >= 0) |
133 | return 0; | |
cbdabc7f AA |
134 | |
135 | switch (task->tk_status) { | |
9cb81968 | 136 | /* DS session errors */ |
cbdabc7f AA |
137 | case -NFS4ERR_BADSESSION: |
138 | case -NFS4ERR_BADSLOT: | |
139 | case -NFS4ERR_BAD_HIGH_SLOT: | |
140 | case -NFS4ERR_DEADSESSION: | |
141 | case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION: | |
142 | case -NFS4ERR_SEQ_FALSE_RETRY: | |
143 | case -NFS4ERR_SEQ_MISORDERED: | |
144 | dprintk("%s ERROR %d, Reset session. Exchangeid " | |
145 | "flags 0x%x\n", __func__, task->tk_status, | |
146 | clp->cl_exchange_flags); | |
9f594791 | 147 | nfs4_schedule_session_recovery(clp->cl_session, task->tk_status); |
cbdabc7f AA |
148 | break; |
149 | case -NFS4ERR_DELAY: | |
150 | case -NFS4ERR_GRACE: | |
cbdabc7f AA |
151 | rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX); |
152 | break; | |
a8a4ae3a AA |
153 | case -NFS4ERR_RETRY_UNCACHED_REP: |
154 | break; | |
041245c8 | 155 | /* Invalidate Layout errors */ |
a0bc01e0 | 156 | case -NFS4ERR_ACCESS: |
041245c8 AA |
157 | case -NFS4ERR_PNFS_NO_LAYOUT: |
158 | case -ESTALE: /* mapped NFS4ERR_STALE */ | |
159 | case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */ | |
160 | case -EISDIR: /* mapped NFS4ERR_ISDIR */ | |
161 | case -NFS4ERR_FHEXPIRED: | |
162 | case -NFS4ERR_WRONG_TYPE: | |
163 | dprintk("%s Invalid layout error %d\n", __func__, | |
164 | task->tk_status); | |
165 | /* | |
166 | * Destroy layout so new i/o will get a new layout. | |
167 | * Layout will not be destroyed until all current lseg | |
168 | * references are put. Mark layout as invalid to resend failed | |
169 | * i/o and all i/o waiting on the slot table to the MDS until | |
170 | * layout is destroyed and a new valid layout is obtained. | |
171 | */ | |
d42e7873 | 172 | pnfs_destroy_layout(NFS_I(inode)); |
041245c8 AA |
173 | rpc_wake_up(&tbl->slot_tbl_waitq); |
174 | goto reset; | |
e7dd79af AA |
175 | /* RPC connection errors */ |
176 | case -ECONNREFUSED: | |
177 | case -EHOSTDOWN: | |
178 | case -EHOSTUNREACH: | |
179 | case -ENETUNREACH: | |
180 | case -EIO: | |
181 | case -ETIMEDOUT: | |
182 | case -EPIPE: | |
183 | dprintk("%s DS connection error %d\n", __func__, | |
184 | task->tk_status); | |
1dfed273 | 185 | nfs4_mark_deviceid_unavailable(devid); |
c220106f | 186 | pnfs_error_mark_layout_for_return(inode, lseg); |
a0bc01e0 | 187 | pnfs_set_lo_fail(lseg); |
671fb896 | 188 | rpc_wake_up(&tbl->slot_tbl_waitq); |
e7dd79af | 189 | /* fall through */ |
cbdabc7f | 190 | default: |
041245c8 | 191 | reset: |
e7dd79af | 192 | dprintk("%s Retry through MDS. Error %d\n", __func__, |
cbdabc7f | 193 | task->tk_status); |
e7dd79af | 194 | return -NFS4ERR_RESET_TO_MDS; |
cbdabc7f AA |
195 | } |
196 | task->tk_status = 0; | |
197 | return -EAGAIN; | |
198 | } | |
199 | ||
200 | /* NFS_PROTO call done callback routines */ | |
201 | ||
202 | static int filelayout_read_done_cb(struct rpc_task *task, | |
d45f60c6 | 203 | struct nfs_pgio_header *hdr) |
cbdabc7f | 204 | { |
e7dd79af | 205 | int err; |
cbdabc7f | 206 | |
d45f60c6 WAA |
207 | trace_nfs4_pnfs_read(hdr, task->tk_status); |
208 | err = filelayout_async_handle_error(task, hdr->args.context->state, | |
209 | hdr->ds_clp, hdr->lseg); | |
cbdabc7f | 210 | |
e7dd79af AA |
211 | switch (err) { |
212 | case -NFS4ERR_RESET_TO_MDS: | |
d45f60c6 | 213 | filelayout_reset_read(hdr); |
e7dd79af AA |
214 | return task->tk_status; |
215 | case -EAGAIN: | |
d00c5d43 | 216 | rpc_restart_call_prepare(task); |
cbdabc7f AA |
217 | return -EAGAIN; |
218 | } | |
219 | ||
220 | return 0; | |
221 | } | |
222 | ||
863a3c6c AA |
223 | /* |
224 | * We reference the rpc_cred of the first WRITE that triggers the need for | |
225 | * a LAYOUTCOMMIT, and use it to send the layoutcommit compound. | |
226 | * rfc5661 is not clear about which credential should be used. | |
227 | */ | |
228 | static void | |
d45f60c6 | 229 | filelayout_set_layoutcommit(struct nfs_pgio_header *hdr) |
863a3c6c | 230 | { |
2e18d4d8 | 231 | loff_t end_offs = 0; |
cd841605 FI |
232 | |
233 | if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds || | |
2e18d4d8 | 234 | hdr->res.verf->committed == NFS_FILE_SYNC) |
863a3c6c | 235 | return; |
2e18d4d8 TM |
236 | if (hdr->res.verf->committed == NFS_DATA_SYNC) |
237 | end_offs = hdr->mds_offset + (loff_t)hdr->res.count; | |
863a3c6c | 238 | |
2e18d4d8 TM |
239 | /* Note: if the write is unstable, don't set end_offs until commit */ |
240 | pnfs_set_layoutcommit(hdr->inode, hdr->lseg, end_offs); | |
f383b7e8 | 241 | dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino, |
cd841605 | 242 | (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb); |
863a3c6c AA |
243 | } |
244 | ||
1dfed273 TM |
245 | bool |
246 | filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node) | |
247 | { | |
248 | return filelayout_test_devid_invalid(node) || | |
249 | nfs4_test_deviceid_unavailable(node); | |
250 | } | |
251 | ||
252 | static bool | |
253 | filelayout_reset_to_mds(struct pnfs_layout_segment *lseg) | |
254 | { | |
255 | struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg); | |
256 | ||
8006bfba | 257 | return filelayout_test_devid_unavailable(node); |
1dfed273 TM |
258 | } |
259 | ||
dc70d7b3 AA |
260 | /* |
261 | * Call ops for the async read/write cases | |
262 | * In the case of dense layouts, the offset needs to be reset to its | |
263 | * original value. | |
264 | */ | |
265 | static void filelayout_read_prepare(struct rpc_task *task, void *data) | |
266 | { | |
d45f60c6 | 267 | struct nfs_pgio_header *hdr = data; |
dc70d7b3 | 268 | |
d45f60c6 | 269 | if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) { |
c58c8441 TM |
270 | rpc_exit(task, -EIO); |
271 | return; | |
272 | } | |
d45f60c6 | 273 | if (filelayout_reset_to_mds(hdr->lseg)) { |
0ad2f378 | 274 | dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid); |
d45f60c6 | 275 | filelayout_reset_read(hdr); |
0ad2f378 AA |
276 | rpc_exit(task, 0); |
277 | return; | |
278 | } | |
d45f60c6 | 279 | hdr->pgio_done_cb = filelayout_read_done_cb; |
cbdabc7f | 280 | |
6de7e12f | 281 | if (nfs4_setup_sequence(hdr->ds_clp, |
d45f60c6 WAA |
282 | &hdr->args.seq_args, |
283 | &hdr->res.seq_res, | |
9b206149 TM |
284 | task)) |
285 | return; | |
d45f60c6 WAA |
286 | if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context, |
287 | hdr->args.lock_context, FMODE_READ) == -EIO) | |
869a9d37 | 288 | rpc_exit(task, -EIO); /* lost lock, terminate I/O */ |
dc70d7b3 AA |
289 | } |
290 | ||
291 | static void filelayout_read_call_done(struct rpc_task *task, void *data) | |
292 | { | |
d45f60c6 | 293 | struct nfs_pgio_header *hdr = data; |
dc70d7b3 AA |
294 | |
295 | dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status); | |
296 | ||
d45f60c6 | 297 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags) && |
f9c96fcc | 298 | task->tk_status == 0) { |
d45f60c6 | 299 | nfs41_sequence_done(task, &hdr->res.seq_res); |
0ad2f378 | 300 | return; |
f9c96fcc | 301 | } |
0ad2f378 | 302 | |
dc70d7b3 | 303 | /* Note this may cause RPC to be resent */ |
d45f60c6 | 304 | hdr->mds_ops->rpc_call_done(task, data); |
dc70d7b3 AA |
305 | } |
306 | ||
0a702195 WAA |
307 | static void filelayout_read_count_stats(struct rpc_task *task, void *data) |
308 | { | |
d45f60c6 | 309 | struct nfs_pgio_header *hdr = data; |
0a702195 | 310 | |
d45f60c6 | 311 | rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics); |
0a702195 WAA |
312 | } |
313 | ||
a69aef14 | 314 | static int filelayout_write_done_cb(struct rpc_task *task, |
d45f60c6 | 315 | struct nfs_pgio_header *hdr) |
a69aef14 | 316 | { |
e7dd79af AA |
317 | int err; |
318 | ||
d45f60c6 WAA |
319 | trace_nfs4_pnfs_write(hdr, task->tk_status); |
320 | err = filelayout_async_handle_error(task, hdr->args.context->state, | |
321 | hdr->ds_clp, hdr->lseg); | |
e7dd79af AA |
322 | |
323 | switch (err) { | |
324 | case -NFS4ERR_RESET_TO_MDS: | |
d45f60c6 | 325 | filelayout_reset_write(hdr); |
e7dd79af AA |
326 | return task->tk_status; |
327 | case -EAGAIN: | |
d00c5d43 | 328 | rpc_restart_call_prepare(task); |
a69aef14 FI |
329 | return -EAGAIN; |
330 | } | |
331 | ||
d45f60c6 | 332 | filelayout_set_layoutcommit(hdr); |
e033fb51 TM |
333 | |
334 | /* zero out the fattr */ | |
335 | hdr->fattr.valid = 0; | |
336 | if (task->tk_status >= 0) | |
337 | nfs_writeback_update_inode(hdr); | |
338 | ||
a69aef14 FI |
339 | return 0; |
340 | } | |
341 | ||
e0c2b380 | 342 | static int filelayout_commit_done_cb(struct rpc_task *task, |
0b7c0153 | 343 | struct nfs_commit_data *data) |
e0c2b380 | 344 | { |
e7dd79af AA |
345 | int err; |
346 | ||
cc668ab3 | 347 | trace_nfs4_pnfs_commit_ds(data, task->tk_status); |
e7dd79af AA |
348 | err = filelayout_async_handle_error(task, NULL, data->ds_clp, |
349 | data->lseg); | |
350 | ||
351 | switch (err) { | |
352 | case -NFS4ERR_RESET_TO_MDS: | |
f54bcf2e | 353 | pnfs_generic_prepare_to_resend_writes(data); |
e7dd79af AA |
354 | return -EAGAIN; |
355 | case -EAGAIN: | |
356 | rpc_restart_call_prepare(task); | |
e0c2b380 FI |
357 | return -EAGAIN; |
358 | } | |
359 | ||
73e6c5d8 | 360 | pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb); |
bc7d4b8f | 361 | |
e0c2b380 FI |
362 | return 0; |
363 | } | |
364 | ||
a69aef14 FI |
365 | static void filelayout_write_prepare(struct rpc_task *task, void *data) |
366 | { | |
d45f60c6 | 367 | struct nfs_pgio_header *hdr = data; |
a69aef14 | 368 | |
d45f60c6 | 369 | if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) { |
c58c8441 TM |
370 | rpc_exit(task, -EIO); |
371 | return; | |
372 | } | |
d45f60c6 | 373 | if (filelayout_reset_to_mds(hdr->lseg)) { |
0ad2f378 | 374 | dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid); |
d45f60c6 | 375 | filelayout_reset_write(hdr); |
0ad2f378 AA |
376 | rpc_exit(task, 0); |
377 | return; | |
378 | } | |
6de7e12f | 379 | if (nfs4_setup_sequence(hdr->ds_clp, |
d45f60c6 WAA |
380 | &hdr->args.seq_args, |
381 | &hdr->res.seq_res, | |
9b206149 TM |
382 | task)) |
383 | return; | |
d45f60c6 WAA |
384 | if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context, |
385 | hdr->args.lock_context, FMODE_WRITE) == -EIO) | |
869a9d37 | 386 | rpc_exit(task, -EIO); /* lost lock, terminate I/O */ |
a69aef14 FI |
387 | } |
388 | ||
389 | static void filelayout_write_call_done(struct rpc_task *task, void *data) | |
390 | { | |
d45f60c6 | 391 | struct nfs_pgio_header *hdr = data; |
a69aef14 | 392 | |
d45f60c6 | 393 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags) && |
f9c96fcc | 394 | task->tk_status == 0) { |
d45f60c6 | 395 | nfs41_sequence_done(task, &hdr->res.seq_res); |
0ad2f378 | 396 | return; |
f9c96fcc | 397 | } |
0ad2f378 | 398 | |
a69aef14 | 399 | /* Note this may cause RPC to be resent */ |
d45f60c6 | 400 | hdr->mds_ops->rpc_call_done(task, data); |
a69aef14 FI |
401 | } |
402 | ||
0a702195 WAA |
403 | static void filelayout_write_count_stats(struct rpc_task *task, void *data) |
404 | { | |
d45f60c6 | 405 | struct nfs_pgio_header *hdr = data; |
0a702195 | 406 | |
d45f60c6 | 407 | rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics); |
0a702195 WAA |
408 | } |
409 | ||
0b7c0153 | 410 | static void filelayout_commit_prepare(struct rpc_task *task, void *data) |
e0c2b380 | 411 | { |
0b7c0153 | 412 | struct nfs_commit_data *wdata = data; |
e0c2b380 | 413 | |
6de7e12f | 414 | nfs4_setup_sequence(wdata->ds_clp, |
d9afbd1b TM |
415 | &wdata->args.seq_args, |
416 | &wdata->res.seq_res, | |
417 | task); | |
0b7c0153 FI |
418 | } |
419 | ||
0b7c0153 FI |
420 | static void filelayout_commit_count_stats(struct rpc_task *task, void *data) |
421 | { | |
422 | struct nfs_commit_data *cdata = data; | |
423 | ||
424 | rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics); | |
425 | } | |
426 | ||
17280175 | 427 | static const struct rpc_call_ops filelayout_read_call_ops = { |
dc70d7b3 AA |
428 | .rpc_call_prepare = filelayout_read_prepare, |
429 | .rpc_call_done = filelayout_read_call_done, | |
0a702195 | 430 | .rpc_count_stats = filelayout_read_count_stats, |
f54bcf2e | 431 | .rpc_release = pnfs_generic_rw_release, |
dc70d7b3 AA |
432 | }; |
433 | ||
17280175 | 434 | static const struct rpc_call_ops filelayout_write_call_ops = { |
a69aef14 FI |
435 | .rpc_call_prepare = filelayout_write_prepare, |
436 | .rpc_call_done = filelayout_write_call_done, | |
0a702195 | 437 | .rpc_count_stats = filelayout_write_count_stats, |
f54bcf2e | 438 | .rpc_release = pnfs_generic_rw_release, |
a69aef14 FI |
439 | }; |
440 | ||
17280175 | 441 | static const struct rpc_call_ops filelayout_commit_call_ops = { |
0b7c0153 | 442 | .rpc_call_prepare = filelayout_commit_prepare, |
f54bcf2e | 443 | .rpc_call_done = pnfs_generic_write_commit_done, |
0b7c0153 | 444 | .rpc_count_stats = filelayout_commit_count_stats, |
f54bcf2e | 445 | .rpc_release = pnfs_generic_commit_release, |
e0c2b380 FI |
446 | }; |
447 | ||
dc70d7b3 | 448 | static enum pnfs_try_status |
d45f60c6 | 449 | filelayout_read_pagelist(struct nfs_pgio_header *hdr) |
dc70d7b3 | 450 | { |
cd841605 | 451 | struct pnfs_layout_segment *lseg = hdr->lseg; |
dc70d7b3 | 452 | struct nfs4_pnfs_ds *ds; |
0e20162e | 453 | struct rpc_clnt *ds_clnt; |
d45f60c6 | 454 | loff_t offset = hdr->args.offset; |
dc70d7b3 AA |
455 | u32 j, idx; |
456 | struct nfs_fh *fh; | |
dc70d7b3 | 457 | |
5b5e0928 | 458 | dprintk("--> %s ino %lu pgbase %u req %zu@%llu\n", |
cd841605 | 459 | __func__, hdr->inode->i_ino, |
d45f60c6 | 460 | hdr->args.pgbase, (size_t)hdr->args.count, offset); |
dc70d7b3 AA |
461 | |
462 | /* Retrieve the correct rpc_client for the byte range */ | |
463 | j = nfs4_fl_calc_j_index(lseg, offset); | |
464 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
465 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 466 | if (!ds) |
dc70d7b3 | 467 | return PNFS_NOT_ATTEMPTED; |
0e20162e AA |
468 | |
469 | ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode); | |
470 | if (IS_ERR(ds_clnt)) | |
471 | return PNFS_NOT_ATTEMPTED; | |
472 | ||
3a7936c3 | 473 | dprintk("%s USE DS: %s cl_count %d\n", __func__, |
212bf41d | 474 | ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count)); |
dc70d7b3 AA |
475 | |
476 | /* No multipath support. Use first DS */ | |
212bf41d | 477 | refcount_inc(&ds->ds_clp->cl_count); |
d45f60c6 | 478 | hdr->ds_clp = ds->ds_clp; |
6cccbb6f | 479 | hdr->ds_commit_idx = idx; |
dc70d7b3 AA |
480 | fh = nfs4_fl_select_ds_fh(lseg, j); |
481 | if (fh) | |
d45f60c6 | 482 | hdr->args.fh = fh; |
dc70d7b3 | 483 | |
d45f60c6 WAA |
484 | hdr->args.offset = filelayout_get_dserver_offset(lseg, offset); |
485 | hdr->mds_offset = offset; | |
dc70d7b3 AA |
486 | |
487 | /* Perform an asynchronous read to ds */ | |
46a5ab47 PT |
488 | nfs_initiate_pgio(ds_clnt, hdr, hdr->cred, |
489 | NFS_PROTO(hdr->inode), &filelayout_read_call_ops, | |
490 | 0, RPC_TASK_SOFTCONN); | |
dc70d7b3 AA |
491 | return PNFS_ATTEMPTED; |
492 | } | |
493 | ||
a69aef14 | 494 | /* Perform async writes. */ |
0382b744 | 495 | static enum pnfs_try_status |
d45f60c6 | 496 | filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync) |
0382b744 | 497 | { |
cd841605 | 498 | struct pnfs_layout_segment *lseg = hdr->lseg; |
a69aef14 | 499 | struct nfs4_pnfs_ds *ds; |
0e20162e | 500 | struct rpc_clnt *ds_clnt; |
d45f60c6 | 501 | loff_t offset = hdr->args.offset; |
a69aef14 FI |
502 | u32 j, idx; |
503 | struct nfs_fh *fh; | |
a69aef14 FI |
504 | |
505 | /* Retrieve the correct rpc_client for the byte range */ | |
506 | j = nfs4_fl_calc_j_index(lseg, offset); | |
507 | idx = nfs4_fl_calc_ds_index(lseg, j); | |
508 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
90fecfcb | 509 | if (!ds) |
a69aef14 | 510 | return PNFS_NOT_ATTEMPTED; |
0e20162e AA |
511 | |
512 | ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode); | |
513 | if (IS_ERR(ds_clnt)) | |
514 | return PNFS_NOT_ATTEMPTED; | |
515 | ||
5b5e0928 | 516 | dprintk("%s ino %lu sync %d req %zu@%llu DS: %s cl_count %d\n", |
d45f60c6 | 517 | __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count, |
212bf41d | 518 | offset, ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count)); |
a69aef14 | 519 | |
d45f60c6 | 520 | hdr->pgio_done_cb = filelayout_write_done_cb; |
212bf41d | 521 | refcount_inc(&ds->ds_clp->cl_count); |
d45f60c6 | 522 | hdr->ds_clp = ds->ds_clp; |
6cccbb6f | 523 | hdr->ds_commit_idx = idx; |
a69aef14 FI |
524 | fh = nfs4_fl_select_ds_fh(lseg, j); |
525 | if (fh) | |
d45f60c6 WAA |
526 | hdr->args.fh = fh; |
527 | hdr->args.offset = filelayout_get_dserver_offset(lseg, offset); | |
a69aef14 FI |
528 | |
529 | /* Perform an asynchronous write */ | |
46a5ab47 PT |
530 | nfs_initiate_pgio(ds_clnt, hdr, hdr->cred, |
531 | NFS_PROTO(hdr->inode), &filelayout_write_call_ops, | |
532 | sync, RPC_TASK_SOFTCONN); | |
a69aef14 | 533 | return PNFS_ATTEMPTED; |
0382b744 AA |
534 | } |
535 | ||
8d40b0f1 AA |
536 | static int |
537 | filelayout_check_deviceid(struct pnfs_layout_hdr *lo, | |
538 | struct nfs4_filelayout_segment *fl, | |
539 | gfp_t gfp_flags) | |
540 | { | |
541 | struct nfs4_deviceid_node *d; | |
542 | struct nfs4_file_layout_dsaddr *dsaddr; | |
543 | int status = -EINVAL; | |
544 | ||
1ebf9801 TM |
545 | /* Is the deviceid already set? If so, we're good. */ |
546 | if (fl->dsaddr != NULL) | |
547 | return 0; | |
548 | ||
8d40b0f1 AA |
549 | /* find and reference the deviceid */ |
550 | d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), &fl->deviceid, | |
551 | lo->plh_lc_cred, gfp_flags); | |
552 | if (d == NULL) | |
553 | goto out; | |
554 | ||
555 | dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node); | |
556 | /* Found deviceid is unavailable */ | |
557 | if (filelayout_test_devid_unavailable(&dsaddr->id_node)) | |
558 | goto out_put; | |
559 | ||
8d40b0f1 AA |
560 | if (fl->first_stripe_index >= dsaddr->stripe_count) { |
561 | dprintk("%s Bad first_stripe_index %u\n", | |
562 | __func__, fl->first_stripe_index); | |
563 | goto out_put; | |
564 | } | |
565 | ||
566 | if ((fl->stripe_type == STRIPE_SPARSE && | |
567 | fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) || | |
568 | (fl->stripe_type == STRIPE_DENSE && | |
569 | fl->num_fh != dsaddr->stripe_count)) { | |
570 | dprintk("%s num_fh %u not valid for given packing\n", | |
571 | __func__, fl->num_fh); | |
572 | goto out_put; | |
573 | } | |
574 | status = 0; | |
1ebf9801 TM |
575 | |
576 | /* | |
577 | * Atomic compare and xchange to ensure we don't scribble | |
578 | * over a non-NULL pointer. | |
579 | */ | |
580 | if (cmpxchg(&fl->dsaddr, NULL, dsaddr) != NULL) | |
581 | goto out_put; | |
8d40b0f1 AA |
582 | out: |
583 | return status; | |
584 | out_put: | |
585 | nfs4_fl_put_deviceid(dsaddr); | |
586 | goto out; | |
587 | } | |
588 | ||
16b374ca AA |
589 | /* |
590 | * filelayout_check_layout() | |
591 | * | |
592 | * Make sure layout segment parameters are sane WRT the device. | |
593 | * At this point no generic layer initialization of the lseg has occurred, | |
594 | * and nothing has been added to the layout_hdr cache. | |
595 | * | |
596 | */ | |
597 | static int | |
598 | filelayout_check_layout(struct pnfs_layout_hdr *lo, | |
599 | struct nfs4_filelayout_segment *fl, | |
600 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 | 601 | gfp_t gfp_flags) |
16b374ca | 602 | { |
16b374ca | 603 | int status = -EINVAL; |
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 | ||
c6194271 | 621 | if (!fl->stripe_unit) { |
75247aff | 622 | dprintk("%s Invalid stripe unit (%u)\n", |
16b374ca AA |
623 | __func__, fl->stripe_unit); |
624 | goto out; | |
625 | } | |
626 | ||
16b374ca AA |
627 | status = 0; |
628 | out: | |
629 | dprintk("--> %s returns %d\n", __func__, status); | |
630 | return status; | |
16b374ca AA |
631 | } |
632 | ||
3ec0c979 | 633 | static void _filelayout_free_lseg(struct nfs4_filelayout_segment *fl) |
16b374ca AA |
634 | { |
635 | int i; | |
636 | ||
3ec0c979 KM |
637 | if (fl->fh_array) { |
638 | for (i = 0; i < fl->num_fh; i++) { | |
639 | if (!fl->fh_array[i]) | |
640 | break; | |
641 | kfree(fl->fh_array[i]); | |
642 | } | |
643 | kfree(fl->fh_array); | |
16b374ca | 644 | } |
16b374ca AA |
645 | kfree(fl); |
646 | } | |
647 | ||
648 | static int | |
649 | filelayout_decode_layout(struct pnfs_layout_hdr *flo, | |
650 | struct nfs4_filelayout_segment *fl, | |
651 | struct nfs4_layoutget_res *lgr, | |
a75b9df9 | 652 | gfp_t gfp_flags) |
16b374ca | 653 | { |
35124a09 | 654 | struct xdr_stream stream; |
f7da7a12 | 655 | struct xdr_buf buf; |
35124a09 WAA |
656 | struct page *scratch; |
657 | __be32 *p; | |
16b374ca AA |
658 | uint32_t nfl_util; |
659 | int i; | |
660 | ||
661 | dprintk("%s: set_layout_map Begin\n", __func__); | |
662 | ||
a75b9df9 | 663 | scratch = alloc_page(gfp_flags); |
35124a09 WAA |
664 | if (!scratch) |
665 | return -ENOMEM; | |
666 | ||
f7da7a12 | 667 | xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len); |
35124a09 WAA |
668 | xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE); |
669 | ||
670 | /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8), | |
671 | * num_fh (4) */ | |
672 | p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20); | |
673 | if (unlikely(!p)) | |
674 | goto out_err; | |
675 | ||
629dc870 | 676 | memcpy(&fl->deviceid, p, sizeof(fl->deviceid)); |
16b374ca | 677 | p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE); |
629dc870 | 678 | nfs4_print_deviceid(&fl->deviceid); |
16b374ca AA |
679 | |
680 | nfl_util = be32_to_cpup(p++); | |
681 | if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS) | |
682 | fl->commit_through_mds = 1; | |
683 | if (nfl_util & NFL4_UFLG_DENSE) | |
684 | fl->stripe_type = STRIPE_DENSE; | |
685 | else | |
686 | fl->stripe_type = STRIPE_SPARSE; | |
687 | fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK; | |
688 | ||
689 | fl->first_stripe_index = be32_to_cpup(p++); | |
690 | p = xdr_decode_hyper(p, &fl->pattern_offset); | |
691 | fl->num_fh = be32_to_cpup(p++); | |
692 | ||
693 | dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n", | |
694 | __func__, nfl_util, fl->num_fh, fl->first_stripe_index, | |
695 | fl->pattern_offset); | |
696 | ||
cec765cf AA |
697 | /* Note that a zero value for num_fh is legal for STRIPE_SPARSE. |
698 | * Futher checking is done in filelayout_check_layout */ | |
e414966b | 699 | if (fl->num_fh > |
cec765cf | 700 | max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT)) |
35124a09 WAA |
701 | goto out_err; |
702 | ||
cec765cf | 703 | if (fl->num_fh > 0) { |
1813badd | 704 | fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]), |
cec765cf AA |
705 | gfp_flags); |
706 | if (!fl->fh_array) | |
707 | goto out_err; | |
708 | } | |
16b374ca AA |
709 | |
710 | for (i = 0; i < fl->num_fh; i++) { | |
711 | /* Do we want to use a mempool here? */ | |
a75b9df9 | 712 | fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags); |
35124a09 | 713 | if (!fl->fh_array[i]) |
3ec0c979 | 714 | goto out_err; |
35124a09 WAA |
715 | |
716 | p = xdr_inline_decode(&stream, 4); | |
717 | if (unlikely(!p)) | |
3ec0c979 | 718 | goto out_err; |
16b374ca AA |
719 | fl->fh_array[i]->size = be32_to_cpup(p++); |
720 | if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) { | |
f9fd2d9c | 721 | printk(KERN_ERR "NFS: Too big fh %d received %d\n", |
16b374ca | 722 | i, fl->fh_array[i]->size); |
3ec0c979 | 723 | goto out_err; |
16b374ca | 724 | } |
35124a09 WAA |
725 | |
726 | p = xdr_inline_decode(&stream, fl->fh_array[i]->size); | |
727 | if (unlikely(!p)) | |
3ec0c979 | 728 | goto out_err; |
16b374ca | 729 | memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size); |
16b374ca AA |
730 | dprintk("DEBUG: %s: fh len %d\n", __func__, |
731 | fl->fh_array[i]->size); | |
732 | } | |
733 | ||
35124a09 | 734 | __free_page(scratch); |
16b374ca | 735 | return 0; |
35124a09 | 736 | |
35124a09 WAA |
737 | out_err: |
738 | __free_page(scratch); | |
739 | return -EIO; | |
16b374ca AA |
740 | } |
741 | ||
c879513e FI |
742 | static void |
743 | filelayout_free_lseg(struct pnfs_layout_segment *lseg) | |
744 | { | |
745 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
746 | ||
747 | dprintk("--> %s\n", __func__); | |
0a47df11 SM |
748 | if (fl->dsaddr != NULL) |
749 | nfs4_fl_put_deviceid(fl->dsaddr); | |
799ba8d5 FI |
750 | /* This assumes a single RW lseg */ |
751 | if (lseg->pls_range.iomode == IOMODE_RW) { | |
752 | struct nfs4_filelayout *flo; | |
753 | ||
754 | flo = FILELAYOUT_FROM_HDR(lseg->pls_layout); | |
755 | flo->commit_info.nbuckets = 0; | |
756 | kfree(flo->commit_info.buckets); | |
757 | flo->commit_info.buckets = NULL; | |
758 | } | |
c879513e FI |
759 | _filelayout_free_lseg(fl); |
760 | } | |
761 | ||
799ba8d5 FI |
762 | static int |
763 | filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg, | |
ea2cf228 | 764 | struct nfs_commit_info *cinfo, |
799ba8d5 FI |
765 | gfp_t gfp_flags) |
766 | { | |
767 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); | |
ea2cf228 | 768 | struct pnfs_commit_bucket *buckets; |
dd7663e7 | 769 | int size, i; |
799ba8d5 FI |
770 | |
771 | if (fl->commit_through_mds) | |
772 | return 0; | |
dd7663e7 WAA |
773 | |
774 | size = (fl->stripe_type == STRIPE_SPARSE) ? | |
775 | fl->dsaddr->ds_num : fl->dsaddr->stripe_count; | |
776 | ||
777 | if (cinfo->ds->nbuckets >= size) { | |
799ba8d5 FI |
778 | /* This assumes there is only one IOMODE_RW lseg. What |
779 | * we really want to do is have a layout_hdr level | |
780 | * dictionary of <multipath_list4, fh> keys, each | |
781 | * associated with a struct list_head, populated by calls | |
782 | * to filelayout_write_pagelist(). | |
783 | * */ | |
784 | return 0; | |
785 | } | |
786 | ||
ea2cf228 | 787 | buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket), |
799ba8d5 FI |
788 | gfp_flags); |
789 | if (!buckets) | |
790 | return -ENOMEM; | |
dd7663e7 WAA |
791 | for (i = 0; i < size; i++) { |
792 | INIT_LIST_HEAD(&buckets[i].written); | |
793 | INIT_LIST_HEAD(&buckets[i].committing); | |
5002c586 WAA |
794 | /* mark direct verifier as unset */ |
795 | buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW; | |
799ba8d5 | 796 | } |
dd7663e7 | 797 | |
fe238e60 | 798 | spin_lock(&cinfo->inode->i_lock); |
dd7663e7 WAA |
799 | if (cinfo->ds->nbuckets >= size) |
800 | goto out; | |
801 | for (i = 0; i < cinfo->ds->nbuckets; i++) { | |
802 | list_splice(&cinfo->ds->buckets[i].written, | |
803 | &buckets[i].written); | |
804 | list_splice(&cinfo->ds->buckets[i].committing, | |
805 | &buckets[i].committing); | |
5002c586 WAA |
806 | buckets[i].direct_verf.committed = |
807 | cinfo->ds->buckets[i].direct_verf.committed; | |
dd7663e7 WAA |
808 | buckets[i].wlseg = cinfo->ds->buckets[i].wlseg; |
809 | buckets[i].clseg = cinfo->ds->buckets[i].clseg; | |
810 | } | |
811 | swap(cinfo->ds->buckets, buckets); | |
812 | cinfo->ds->nbuckets = size; | |
813 | out: | |
fe238e60 | 814 | spin_unlock(&cinfo->inode->i_lock); |
dd7663e7 WAA |
815 | kfree(buckets); |
816 | return 0; | |
799ba8d5 FI |
817 | } |
818 | ||
16b374ca AA |
819 | static struct pnfs_layout_segment * |
820 | filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid, | |
a75b9df9 TM |
821 | struct nfs4_layoutget_res *lgr, |
822 | gfp_t gfp_flags) | |
16b374ca AA |
823 | { |
824 | struct nfs4_filelayout_segment *fl; | |
825 | int rc; | |
16b374ca AA |
826 | |
827 | dprintk("--> %s\n", __func__); | |
a75b9df9 | 828 | fl = kzalloc(sizeof(*fl), gfp_flags); |
16b374ca AA |
829 | if (!fl) |
830 | return NULL; | |
831 | ||
629dc870 AA |
832 | rc = filelayout_decode_layout(layoutid, fl, lgr, gfp_flags); |
833 | if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, gfp_flags)) { | |
16b374ca AA |
834 | _filelayout_free_lseg(fl); |
835 | return NULL; | |
836 | } | |
837 | return &fl->generic_hdr; | |
838 | } | |
839 | ||
94ad1c80 FI |
840 | /* |
841 | * filelayout_pg_test(). Called by nfs_can_coalesce_requests() | |
842 | * | |
b4fdac1a WAA |
843 | * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number |
844 | * of bytes (maximum @req->wb_bytes) that can be coalesced. | |
94ad1c80 | 845 | */ |
b4fdac1a | 846 | static size_t |
94ad1c80 FI |
847 | filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev, |
848 | struct nfs_page *req) | |
849 | { | |
0f9c429e | 850 | unsigned int size; |
94ad1c80 | 851 | u64 p_stripe, r_stripe; |
c6194271 WAA |
852 | u32 stripe_offset; |
853 | u64 segment_offset = pgio->pg_lseg->pls_range.offset; | |
854 | u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit; | |
94ad1c80 | 855 | |
0f9c429e WAA |
856 | /* calls nfs_generic_pg_test */ |
857 | size = pnfs_generic_pg_test(pgio, prev, req); | |
858 | if (!size) | |
b4fdac1a | 859 | return 0; |
89a58e32 | 860 | |
c6194271 | 861 | /* see if req and prev are in the same stripe */ |
0f9c429e | 862 | if (prev) { |
c6194271 WAA |
863 | p_stripe = (u64)req_offset(prev) - segment_offset; |
864 | r_stripe = (u64)req_offset(req) - segment_offset; | |
0f9c429e WAA |
865 | do_div(p_stripe, stripe_unit); |
866 | do_div(r_stripe, stripe_unit); | |
94ad1c80 | 867 | |
0f9c429e WAA |
868 | if (p_stripe != r_stripe) |
869 | return 0; | |
870 | } | |
c6194271 WAA |
871 | |
872 | /* calculate remaining bytes in the current stripe */ | |
873 | div_u64_rem((u64)req_offset(req) - segment_offset, | |
874 | stripe_unit, | |
875 | &stripe_offset); | |
876 | WARN_ON_ONCE(stripe_offset > stripe_unit); | |
877 | if (stripe_offset >= stripe_unit) | |
878 | return 0; | |
879 | return min(stripe_unit - (unsigned int)stripe_offset, size); | |
94ad1c80 FI |
880 | } |
881 | ||
8d40b0f1 AA |
882 | static struct pnfs_layout_segment * |
883 | fl_pnfs_update_layout(struct inode *ino, | |
884 | struct nfs_open_context *ctx, | |
885 | loff_t pos, | |
886 | u64 count, | |
887 | enum pnfs_iomode iomode, | |
888 | bool strict_iomode, | |
889 | gfp_t gfp_flags) | |
890 | { | |
891 | struct pnfs_layout_segment *lseg = NULL; | |
892 | struct pnfs_layout_hdr *lo; | |
893 | struct nfs4_filelayout_segment *fl; | |
894 | int status; | |
895 | ||
896 | lseg = pnfs_update_layout(ino, ctx, pos, count, iomode, strict_iomode, | |
897 | gfp_flags); | |
1935cf5d | 898 | if (IS_ERR_OR_NULL(lseg)) |
8d40b0f1 AA |
899 | goto out; |
900 | ||
901 | lo = NFS_I(ino)->layout; | |
902 | fl = FILELAYOUT_LSEG(lseg); | |
903 | ||
904 | status = filelayout_check_deviceid(lo, fl, gfp_flags); | |
209aa230 AS |
905 | if (status) { |
906 | pnfs_put_lseg(lseg); | |
8d40b0f1 | 907 | lseg = ERR_PTR(status); |
209aa230 | 908 | } |
8d40b0f1 | 909 | out: |
8d40b0f1 AA |
910 | return lseg; |
911 | } | |
912 | ||
17280175 | 913 | static void |
7c24d948 AA |
914 | filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio, |
915 | struct nfs_page *req) | |
916 | { | |
b3230e80 | 917 | pnfs_generic_pg_check_layout(pgio); |
d600ad1f | 918 | if (!pgio->pg_lseg) { |
8d40b0f1 AA |
919 | pgio->pg_lseg = fl_pnfs_update_layout(pgio->pg_inode, |
920 | req->wb_context, | |
921 | 0, | |
922 | NFS4_MAX_UINT64, | |
923 | IOMODE_READ, | |
924 | false, | |
925 | GFP_KERNEL); | |
d600ad1f PT |
926 | if (IS_ERR(pgio->pg_lseg)) { |
927 | pgio->pg_error = PTR_ERR(pgio->pg_lseg); | |
928 | pgio->pg_lseg = NULL; | |
929 | return; | |
930 | } | |
931 | } | |
7c24d948 AA |
932 | /* If no lseg, fall back to read through mds */ |
933 | if (pgio->pg_lseg == NULL) | |
1f945357 | 934 | nfs_pageio_reset_read_mds(pgio); |
7c24d948 AA |
935 | } |
936 | ||
17280175 | 937 | static void |
7c24d948 AA |
938 | filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio, |
939 | struct nfs_page *req) | |
940 | { | |
ea2cf228 | 941 | struct nfs_commit_info cinfo; |
799ba8d5 FI |
942 | int status; |
943 | ||
b3230e80 | 944 | pnfs_generic_pg_check_layout(pgio); |
d600ad1f | 945 | if (!pgio->pg_lseg) { |
8d40b0f1 AA |
946 | pgio->pg_lseg = fl_pnfs_update_layout(pgio->pg_inode, |
947 | req->wb_context, | |
948 | 0, | |
949 | NFS4_MAX_UINT64, | |
950 | IOMODE_RW, | |
951 | false, | |
952 | GFP_NOFS); | |
d600ad1f PT |
953 | if (IS_ERR(pgio->pg_lseg)) { |
954 | pgio->pg_error = PTR_ERR(pgio->pg_lseg); | |
955 | pgio->pg_lseg = NULL; | |
956 | return; | |
957 | } | |
958 | } | |
959 | ||
7c24d948 AA |
960 | /* If no lseg, fall back to write through mds */ |
961 | if (pgio->pg_lseg == NULL) | |
799ba8d5 | 962 | goto out_mds; |
ea2cf228 FI |
963 | nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq); |
964 | status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS); | |
799ba8d5 | 965 | if (status < 0) { |
9369a431 | 966 | pnfs_put_lseg(pgio->pg_lseg); |
799ba8d5 FI |
967 | pgio->pg_lseg = NULL; |
968 | goto out_mds; | |
969 | } | |
970 | return; | |
971 | out_mds: | |
972 | nfs_pageio_reset_write_mds(pgio); | |
7c24d948 AA |
973 | } |
974 | ||
1751c363 | 975 | static const struct nfs_pageio_ops filelayout_pg_read_ops = { |
7c24d948 | 976 | .pg_init = filelayout_pg_init_read, |
1751c363 | 977 | .pg_test = filelayout_pg_test, |
493292dd | 978 | .pg_doio = pnfs_generic_pg_readpages, |
180bb5ec | 979 | .pg_cleanup = pnfs_generic_pg_cleanup, |
1751c363 TM |
980 | }; |
981 | ||
982 | static const struct nfs_pageio_ops filelayout_pg_write_ops = { | |
7c24d948 | 983 | .pg_init = filelayout_pg_init_write, |
1751c363 | 984 | .pg_test = filelayout_pg_test, |
dce81290 | 985 | .pg_doio = pnfs_generic_pg_writepages, |
180bb5ec | 986 | .pg_cleanup = pnfs_generic_pg_cleanup, |
1751c363 TM |
987 | }; |
988 | ||
e0c2b380 FI |
989 | static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j) |
990 | { | |
991 | if (fl->stripe_type == STRIPE_SPARSE) | |
992 | return nfs4_fl_calc_ds_index(&fl->generic_hdr, j); | |
993 | else | |
994 | return j; | |
995 | } | |
996 | ||
c8a3292d PT |
997 | static void |
998 | filelayout_mark_request_commit(struct nfs_page *req, | |
999 | struct pnfs_layout_segment *lseg, | |
b57ff130 WAA |
1000 | struct nfs_commit_info *cinfo, |
1001 | u32 ds_commit_idx) | |
c8a3292d | 1002 | |
e0c2b380 | 1003 | { |
e0c2b380 FI |
1004 | struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg); |
1005 | u32 i, j; | |
e0c2b380 | 1006 | |
c8a3292d | 1007 | if (fl->commit_through_mds) { |
6272dcc6 | 1008 | nfs_request_add_commit_list(req, cinfo); |
338d00cf TH |
1009 | } else { |
1010 | /* Note that we are calling nfs4_fl_calc_j_index on each page | |
1011 | * that ends up being committed to a data server. An attractive | |
1012 | * alternative is to add a field to nfs_write_data and nfs_page | |
1013 | * to store the value calculated in filelayout_write_pagelist | |
1014 | * and just use that here. | |
d6d6dc7c | 1015 | */ |
338d00cf TH |
1016 | j = nfs4_fl_calc_j_index(lseg, req_offset(req)); |
1017 | i = select_bucket_index(fl, j); | |
1018 | pnfs_layout_mark_request_commit(req, lseg, cinfo, i); | |
e0c2b380 | 1019 | } |
8dd37758 TM |
1020 | } |
1021 | ||
e0c2b380 FI |
1022 | static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i) |
1023 | { | |
1024 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1025 | ||
1026 | if (flseg->stripe_type == STRIPE_SPARSE) | |
1027 | return i; | |
1028 | else | |
1029 | return nfs4_fl_calc_ds_index(lseg, i); | |
1030 | } | |
1031 | ||
1032 | static struct nfs_fh * | |
1033 | select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i) | |
1034 | { | |
1035 | struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg); | |
1036 | ||
1037 | if (flseg->stripe_type == STRIPE_SPARSE) { | |
1038 | if (flseg->num_fh == 1) | |
1039 | i = 0; | |
1040 | else if (flseg->num_fh == 0) | |
1041 | /* Use the MDS OPEN fh set in nfs_read_rpcsetup */ | |
1042 | return NULL; | |
1043 | } | |
1044 | return flseg->fh_array[i]; | |
1045 | } | |
1046 | ||
0b7c0153 | 1047 | static int filelayout_initiate_commit(struct nfs_commit_data *data, int how) |
e0c2b380 FI |
1048 | { |
1049 | struct pnfs_layout_segment *lseg = data->lseg; | |
1050 | struct nfs4_pnfs_ds *ds; | |
0e20162e | 1051 | struct rpc_clnt *ds_clnt; |
e0c2b380 FI |
1052 | u32 idx; |
1053 | struct nfs_fh *fh; | |
1054 | ||
1055 | idx = calc_ds_index_from_commit(lseg, data->ds_commit_index); | |
1056 | ds = nfs4_fl_prepare_ds(lseg, idx); | |
0e20162e AA |
1057 | if (!ds) |
1058 | goto out_err; | |
1059 | ||
1060 | ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode); | |
1061 | if (IS_ERR(ds_clnt)) | |
1062 | goto out_err; | |
1063 | ||
3a7936c3 | 1064 | dprintk("%s ino %lu, how %d cl_count %d\n", __func__, |
212bf41d | 1065 | data->inode->i_ino, how, refcount_read(&ds->ds_clp->cl_count)); |
0b7c0153 | 1066 | data->commit_done_cb = filelayout_commit_done_cb; |
212bf41d | 1067 | refcount_inc(&ds->ds_clp->cl_count); |
e0c2b380 FI |
1068 | data->ds_clp = ds->ds_clp; |
1069 | fh = select_ds_fh_from_commit(lseg, data->ds_commit_index); | |
1070 | if (fh) | |
1071 | data->args.fh = fh; | |
c36aae9a | 1072 | return nfs_initiate_commit(ds_clnt, data, NFS_PROTO(data->inode), |
9f0ec176 AA |
1073 | &filelayout_commit_call_ops, how, |
1074 | RPC_TASK_SOFTCONN); | |
0e20162e | 1075 | out_err: |
f54bcf2e TH |
1076 | pnfs_generic_prepare_to_resend_writes(data); |
1077 | pnfs_generic_commit_release(data); | |
0e20162e | 1078 | return -EAGAIN; |
e0c2b380 FI |
1079 | } |
1080 | ||
02d1426c WAA |
1081 | /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest |
1082 | * for @page | |
1083 | * @cinfo - commit info for current inode | |
1084 | * @page - page to search for matching head request | |
1085 | * | |
1086 | * Returns a the head request if one is found, otherwise returns NULL. | |
1087 | */ | |
1088 | static struct nfs_page * | |
1089 | filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page) | |
1090 | { | |
1091 | struct nfs_page *freq, *t; | |
1092 | struct pnfs_commit_bucket *b; | |
1093 | int i; | |
1094 | ||
1095 | /* Linearly search the commit lists for each bucket until a matching | |
1096 | * request is found */ | |
1097 | for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) { | |
1098 | list_for_each_entry_safe(freq, t, &b->written, wb_list) { | |
1099 | if (freq->wb_page == page) | |
1100 | return freq->wb_head; | |
1101 | } | |
1102 | list_for_each_entry_safe(freq, t, &b->committing, wb_list) { | |
1103 | if (freq->wb_page == page) | |
1104 | return freq->wb_head; | |
1105 | } | |
1106 | } | |
1107 | ||
1108 | return NULL; | |
1109 | } | |
1110 | ||
e0c2b380 FI |
1111 | static int |
1112 | filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages, | |
ea2cf228 | 1113 | int how, struct nfs_commit_info *cinfo) |
e0c2b380 | 1114 | { |
f54bcf2e TH |
1115 | return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo, |
1116 | filelayout_initiate_commit); | |
e0c2b380 | 1117 | } |
f54bcf2e | 1118 | |
661373b1 CH |
1119 | static struct nfs4_deviceid_node * |
1120 | filelayout_alloc_deviceid_node(struct nfs_server *server, | |
1121 | struct pnfs_device *pdev, gfp_t gfp_flags) | |
1122 | { | |
1123 | struct nfs4_file_layout_dsaddr *dsaddr; | |
1124 | ||
1125 | dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags); | |
1126 | if (!dsaddr) | |
1127 | return NULL; | |
1128 | return &dsaddr->id_node; | |
1129 | } | |
e0c2b380 | 1130 | |
1775bc34 | 1131 | static void |
fc87701b | 1132 | filelayout_free_deviceid_node(struct nfs4_deviceid_node *d) |
1775bc34 BH |
1133 | { |
1134 | nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node)); | |
1135 | } | |
1136 | ||
799ba8d5 FI |
1137 | static struct pnfs_layout_hdr * |
1138 | filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags) | |
1139 | { | |
1140 | struct nfs4_filelayout *flo; | |
1141 | ||
1142 | flo = kzalloc(sizeof(*flo), gfp_flags); | |
6df200f5 | 1143 | return flo != NULL ? &flo->generic_hdr : NULL; |
799ba8d5 FI |
1144 | } |
1145 | ||
1146 | static void | |
1147 | filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo) | |
1148 | { | |
1149 | kfree(FILELAYOUT_FROM_HDR(lo)); | |
1150 | } | |
1151 | ||
ea2cf228 FI |
1152 | static struct pnfs_ds_commit_info * |
1153 | filelayout_get_ds_info(struct inode *inode) | |
1154 | { | |
df011748 FI |
1155 | struct pnfs_layout_hdr *layout = NFS_I(inode)->layout; |
1156 | ||
1157 | if (layout == NULL) | |
1158 | return NULL; | |
1159 | else | |
1160 | return &FILELAYOUT_FROM_HDR(layout)->commit_info; | |
ea2cf228 FI |
1161 | } |
1162 | ||
7ab672ce | 1163 | static struct pnfs_layoutdriver_type filelayout_type = { |
ea8eecdd CH |
1164 | .id = LAYOUT_NFSV4_1_FILES, |
1165 | .name = "LAYOUT_NFSV4_1_FILES", | |
1166 | .owner = THIS_MODULE, | |
799ba8d5 FI |
1167 | .alloc_layout_hdr = filelayout_alloc_layout_hdr, |
1168 | .free_layout_hdr = filelayout_free_layout_hdr, | |
ea8eecdd CH |
1169 | .alloc_lseg = filelayout_alloc_lseg, |
1170 | .free_lseg = filelayout_free_lseg, | |
1751c363 TM |
1171 | .pg_read_ops = &filelayout_pg_read_ops, |
1172 | .pg_write_ops = &filelayout_pg_write_ops, | |
ea2cf228 | 1173 | .get_ds_info = &filelayout_get_ds_info, |
8dd37758 | 1174 | .mark_request_commit = filelayout_mark_request_commit, |
f54bcf2e TH |
1175 | .clear_request_commit = pnfs_generic_clear_request_commit, |
1176 | .scan_commit_lists = pnfs_generic_scan_commit_lists, | |
1177 | .recover_commit_reqs = pnfs_generic_recover_commit_reqs, | |
02d1426c | 1178 | .search_commit_reqs = filelayout_search_commit_reqs, |
e0c2b380 | 1179 | .commit_pagelist = filelayout_commit_pagelist, |
dc70d7b3 | 1180 | .read_pagelist = filelayout_read_pagelist, |
0382b744 | 1181 | .write_pagelist = filelayout_write_pagelist, |
661373b1 | 1182 | .alloc_deviceid_node = filelayout_alloc_deviceid_node, |
fc87701b | 1183 | .free_deviceid_node = filelayout_free_deviceid_node, |
5bb89b47 | 1184 | .sync = pnfs_nfs_generic_sync, |
7ab672ce DH |
1185 | }; |
1186 | ||
1187 | static int __init nfs4filelayout_init(void) | |
1188 | { | |
1189 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n", | |
1190 | __func__); | |
1191 | return pnfs_register_layoutdriver(&filelayout_type); | |
1192 | } | |
1193 | ||
1194 | static void __exit nfs4filelayout_exit(void) | |
1195 | { | |
1196 | printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n", | |
1197 | __func__); | |
1198 | pnfs_unregister_layoutdriver(&filelayout_type); | |
1199 | } | |
1200 | ||
f85ef69c BF |
1201 | MODULE_ALIAS("nfs-layouttype4-1"); |
1202 | ||
7ab672ce DH |
1203 | module_init(nfs4filelayout_init); |
1204 | module_exit(nfs4filelayout_exit); |