Merge tag 'gvt-next-2017-02-24' of https://github.com/01org/gvt-linux into drm-intel...
[sfrench/cifs-2.6.git] / fs / nfs / filelayout / filelayout.c
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>
33 #include <linux/nfs_page.h>
34 #include <linux/module.h>
35 #include <linux/backing-dev.h>
36
37 #include <linux/sunrpc/metrics.h>
38
39 #include "../nfs4session.h"
40 #include "../internal.h"
41 #include "../delegation.h"
42 #include "filelayout.h"
43 #include "../nfs4trace.h"
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
51 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
52
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;
58         u64 stripe_no;
59         u32 rem;
60
61         offset -= flseg->pattern_offset;
62         stripe_no = div_u64(offset, stripe_width);
63         div_u64_rem(offset, flseg->stripe_unit, &rem);
64
65         return stripe_no * flseg->stripe_unit + rem;
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
88 static void filelayout_reset_write(struct nfs_pgio_header *hdr)
89 {
90         struct rpc_task *task = &hdr->task;
91
92         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
93                 dprintk("%s Reset task %5u for i/o through MDS "
94                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
95                         hdr->task.tk_pid,
96                         hdr->inode->i_sb->s_id,
97                         (unsigned long long)NFS_FILEID(hdr->inode),
98                         hdr->args.count,
99                         (unsigned long long)hdr->args.offset);
100
101                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
102         }
103 }
104
105 static void filelayout_reset_read(struct nfs_pgio_header *hdr)
106 {
107         struct rpc_task *task = &hdr->task;
108
109         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
110                 dprintk("%s Reset task %5u for i/o through MDS "
111                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
112                         hdr->task.tk_pid,
113                         hdr->inode->i_sb->s_id,
114                         (unsigned long long)NFS_FILEID(hdr->inode),
115                         hdr->args.count,
116                         (unsigned long long)hdr->args.offset);
117
118                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
119         }
120 }
121
122 static int filelayout_async_handle_error(struct rpc_task *task,
123                                          struct nfs4_state *state,
124                                          struct nfs_client *clp,
125                                          struct pnfs_layout_segment *lseg)
126 {
127         struct pnfs_layout_hdr *lo = lseg->pls_layout;
128         struct inode *inode = lo->plh_inode;
129         struct nfs_server *mds_server = NFS_SERVER(inode);
130         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
131         struct nfs_client *mds_client = mds_server->nfs_client;
132         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
133
134         if (task->tk_status >= 0)
135                 return 0;
136
137         switch (task->tk_status) {
138         /* MDS state errors */
139         case -NFS4ERR_DELEG_REVOKED:
140         case -NFS4ERR_ADMIN_REVOKED:
141         case -NFS4ERR_BAD_STATEID:
142         case -NFS4ERR_OPENMODE:
143                 if (state == NULL)
144                         break;
145                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
146                         goto out_bad_stateid;
147                 goto wait_on_recovery;
148         case -NFS4ERR_EXPIRED:
149                 if (state != NULL) {
150                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
151                                 goto out_bad_stateid;
152                 }
153                 nfs4_schedule_lease_recovery(mds_client);
154                 goto wait_on_recovery;
155         /* DS session errors */
156         case -NFS4ERR_BADSESSION:
157         case -NFS4ERR_BADSLOT:
158         case -NFS4ERR_BAD_HIGH_SLOT:
159         case -NFS4ERR_DEADSESSION:
160         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
161         case -NFS4ERR_SEQ_FALSE_RETRY:
162         case -NFS4ERR_SEQ_MISORDERED:
163                 dprintk("%s ERROR %d, Reset session. Exchangeid "
164                         "flags 0x%x\n", __func__, task->tk_status,
165                         clp->cl_exchange_flags);
166                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
167                 break;
168         case -NFS4ERR_DELAY:
169         case -NFS4ERR_GRACE:
170                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
171                 break;
172         case -NFS4ERR_RETRY_UNCACHED_REP:
173                 break;
174         /* Invalidate Layout errors */
175         case -NFS4ERR_PNFS_NO_LAYOUT:
176         case -ESTALE:           /* mapped NFS4ERR_STALE */
177         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
178         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
179         case -NFS4ERR_FHEXPIRED:
180         case -NFS4ERR_WRONG_TYPE:
181                 dprintk("%s Invalid layout error %d\n", __func__,
182                         task->tk_status);
183                 /*
184                  * Destroy layout so new i/o will get a new layout.
185                  * Layout will not be destroyed until all current lseg
186                  * references are put. Mark layout as invalid to resend failed
187                  * i/o and all i/o waiting on the slot table to the MDS until
188                  * layout is destroyed and a new valid layout is obtained.
189                  */
190                 pnfs_destroy_layout(NFS_I(inode));
191                 rpc_wake_up(&tbl->slot_tbl_waitq);
192                 goto reset;
193         /* RPC connection errors */
194         case -ECONNREFUSED:
195         case -EHOSTDOWN:
196         case -EHOSTUNREACH:
197         case -ENETUNREACH:
198         case -EIO:
199         case -ETIMEDOUT:
200         case -EPIPE:
201                 dprintk("%s DS connection error %d\n", __func__,
202                         task->tk_status);
203                 nfs4_mark_deviceid_unavailable(devid);
204                 pnfs_error_mark_layout_for_return(inode, lseg);
205                 pnfs_set_lo_fail(lseg);
206                 rpc_wake_up(&tbl->slot_tbl_waitq);
207                 /* fall through */
208         default:
209 reset:
210                 dprintk("%s Retry through MDS. Error %d\n", __func__,
211                         task->tk_status);
212                 return -NFS4ERR_RESET_TO_MDS;
213         }
214 out:
215         task->tk_status = 0;
216         return -EAGAIN;
217 out_bad_stateid:
218         task->tk_status = -EIO;
219         return 0;
220 wait_on_recovery:
221         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
222         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
223                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
224         goto out;
225 }
226
227 /* NFS_PROTO call done callback routines */
228
229 static int filelayout_read_done_cb(struct rpc_task *task,
230                                 struct nfs_pgio_header *hdr)
231 {
232         int err;
233
234         trace_nfs4_pnfs_read(hdr, task->tk_status);
235         err = filelayout_async_handle_error(task, hdr->args.context->state,
236                                             hdr->ds_clp, hdr->lseg);
237
238         switch (err) {
239         case -NFS4ERR_RESET_TO_MDS:
240                 filelayout_reset_read(hdr);
241                 return task->tk_status;
242         case -EAGAIN:
243                 rpc_restart_call_prepare(task);
244                 return -EAGAIN;
245         }
246
247         return 0;
248 }
249
250 /*
251  * We reference the rpc_cred of the first WRITE that triggers the need for
252  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
253  * rfc5661 is not clear about which credential should be used.
254  */
255 static void
256 filelayout_set_layoutcommit(struct nfs_pgio_header *hdr)
257 {
258         loff_t end_offs = 0;
259
260         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
261             hdr->res.verf->committed == NFS_FILE_SYNC)
262                 return;
263         if (hdr->res.verf->committed == NFS_DATA_SYNC)
264                 end_offs = hdr->mds_offset + (loff_t)hdr->res.count;
265
266         /* Note: if the write is unstable, don't set end_offs until commit */
267         pnfs_set_layoutcommit(hdr->inode, hdr->lseg, end_offs);
268         dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
269                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
270 }
271
272 bool
273 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
274 {
275         return filelayout_test_devid_invalid(node) ||
276                 nfs4_test_deviceid_unavailable(node);
277 }
278
279 static bool
280 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
281 {
282         struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
283
284         return filelayout_test_devid_unavailable(node);
285 }
286
287 /*
288  * Call ops for the async read/write cases
289  * In the case of dense layouts, the offset needs to be reset to its
290  * original value.
291  */
292 static void filelayout_read_prepare(struct rpc_task *task, void *data)
293 {
294         struct nfs_pgio_header *hdr = data;
295
296         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
297                 rpc_exit(task, -EIO);
298                 return;
299         }
300         if (filelayout_reset_to_mds(hdr->lseg)) {
301                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
302                 filelayout_reset_read(hdr);
303                 rpc_exit(task, 0);
304                 return;
305         }
306         hdr->pgio_done_cb = filelayout_read_done_cb;
307
308         if (nfs4_setup_sequence(hdr->ds_clp,
309                         &hdr->args.seq_args,
310                         &hdr->res.seq_res,
311                         task))
312                 return;
313         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
314                         hdr->args.lock_context, FMODE_READ) == -EIO)
315                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
316 }
317
318 static void filelayout_read_call_done(struct rpc_task *task, void *data)
319 {
320         struct nfs_pgio_header *hdr = data;
321
322         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
323
324         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
325             task->tk_status == 0) {
326                 nfs41_sequence_done(task, &hdr->res.seq_res);
327                 return;
328         }
329
330         /* Note this may cause RPC to be resent */
331         hdr->mds_ops->rpc_call_done(task, data);
332 }
333
334 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
335 {
336         struct nfs_pgio_header *hdr = data;
337
338         rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
339 }
340
341 static int filelayout_write_done_cb(struct rpc_task *task,
342                                 struct nfs_pgio_header *hdr)
343 {
344         int err;
345
346         trace_nfs4_pnfs_write(hdr, task->tk_status);
347         err = filelayout_async_handle_error(task, hdr->args.context->state,
348                                             hdr->ds_clp, hdr->lseg);
349
350         switch (err) {
351         case -NFS4ERR_RESET_TO_MDS:
352                 filelayout_reset_write(hdr);
353                 return task->tk_status;
354         case -EAGAIN:
355                 rpc_restart_call_prepare(task);
356                 return -EAGAIN;
357         }
358
359         filelayout_set_layoutcommit(hdr);
360
361         /* zero out the fattr */
362         hdr->fattr.valid = 0;
363         if (task->tk_status >= 0)
364                 nfs_writeback_update_inode(hdr);
365
366         return 0;
367 }
368
369 static int filelayout_commit_done_cb(struct rpc_task *task,
370                                      struct nfs_commit_data *data)
371 {
372         int err;
373
374         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
375         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
376                                             data->lseg);
377
378         switch (err) {
379         case -NFS4ERR_RESET_TO_MDS:
380                 pnfs_generic_prepare_to_resend_writes(data);
381                 return -EAGAIN;
382         case -EAGAIN:
383                 rpc_restart_call_prepare(task);
384                 return -EAGAIN;
385         }
386
387         pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
388
389         return 0;
390 }
391
392 static void filelayout_write_prepare(struct rpc_task *task, void *data)
393 {
394         struct nfs_pgio_header *hdr = data;
395
396         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
397                 rpc_exit(task, -EIO);
398                 return;
399         }
400         if (filelayout_reset_to_mds(hdr->lseg)) {
401                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
402                 filelayout_reset_write(hdr);
403                 rpc_exit(task, 0);
404                 return;
405         }
406         if (nfs4_setup_sequence(hdr->ds_clp,
407                         &hdr->args.seq_args,
408                         &hdr->res.seq_res,
409                         task))
410                 return;
411         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
412                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
413                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
414 }
415
416 static void filelayout_write_call_done(struct rpc_task *task, void *data)
417 {
418         struct nfs_pgio_header *hdr = data;
419
420         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
421             task->tk_status == 0) {
422                 nfs41_sequence_done(task, &hdr->res.seq_res);
423                 return;
424         }
425
426         /* Note this may cause RPC to be resent */
427         hdr->mds_ops->rpc_call_done(task, data);
428 }
429
430 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
431 {
432         struct nfs_pgio_header *hdr = data;
433
434         rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
435 }
436
437 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
438 {
439         struct nfs_commit_data *wdata = data;
440
441         nfs4_setup_sequence(wdata->ds_clp,
442                         &wdata->args.seq_args,
443                         &wdata->res.seq_res,
444                         task);
445 }
446
447 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
448 {
449         struct nfs_commit_data *cdata = data;
450
451         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
452 }
453
454 static const struct rpc_call_ops filelayout_read_call_ops = {
455         .rpc_call_prepare = filelayout_read_prepare,
456         .rpc_call_done = filelayout_read_call_done,
457         .rpc_count_stats = filelayout_read_count_stats,
458         .rpc_release = pnfs_generic_rw_release,
459 };
460
461 static const struct rpc_call_ops filelayout_write_call_ops = {
462         .rpc_call_prepare = filelayout_write_prepare,
463         .rpc_call_done = filelayout_write_call_done,
464         .rpc_count_stats = filelayout_write_count_stats,
465         .rpc_release = pnfs_generic_rw_release,
466 };
467
468 static const struct rpc_call_ops filelayout_commit_call_ops = {
469         .rpc_call_prepare = filelayout_commit_prepare,
470         .rpc_call_done = pnfs_generic_write_commit_done,
471         .rpc_count_stats = filelayout_commit_count_stats,
472         .rpc_release = pnfs_generic_commit_release,
473 };
474
475 static enum pnfs_try_status
476 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
477 {
478         struct pnfs_layout_segment *lseg = hdr->lseg;
479         struct nfs4_pnfs_ds *ds;
480         struct rpc_clnt *ds_clnt;
481         loff_t offset = hdr->args.offset;
482         u32 j, idx;
483         struct nfs_fh *fh;
484
485         dprintk("--> %s ino %lu pgbase %u req %zu@%llu\n",
486                 __func__, hdr->inode->i_ino,
487                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
488
489         /* Retrieve the correct rpc_client for the byte range */
490         j = nfs4_fl_calc_j_index(lseg, offset);
491         idx = nfs4_fl_calc_ds_index(lseg, j);
492         ds = nfs4_fl_prepare_ds(lseg, idx);
493         if (!ds)
494                 return PNFS_NOT_ATTEMPTED;
495
496         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
497         if (IS_ERR(ds_clnt))
498                 return PNFS_NOT_ATTEMPTED;
499
500         dprintk("%s USE DS: %s cl_count %d\n", __func__,
501                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
502
503         /* No multipath support. Use first DS */
504         atomic_inc(&ds->ds_clp->cl_count);
505         hdr->ds_clp = ds->ds_clp;
506         hdr->ds_commit_idx = idx;
507         fh = nfs4_fl_select_ds_fh(lseg, j);
508         if (fh)
509                 hdr->args.fh = fh;
510
511         hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
512         hdr->mds_offset = offset;
513
514         /* Perform an asynchronous read to ds */
515         nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
516                           NFS_PROTO(hdr->inode), &filelayout_read_call_ops,
517                           0, RPC_TASK_SOFTCONN);
518         return PNFS_ATTEMPTED;
519 }
520
521 /* Perform async writes. */
522 static enum pnfs_try_status
523 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
524 {
525         struct pnfs_layout_segment *lseg = hdr->lseg;
526         struct nfs4_pnfs_ds *ds;
527         struct rpc_clnt *ds_clnt;
528         loff_t offset = hdr->args.offset;
529         u32 j, idx;
530         struct nfs_fh *fh;
531
532         /* Retrieve the correct rpc_client for the byte range */
533         j = nfs4_fl_calc_j_index(lseg, offset);
534         idx = nfs4_fl_calc_ds_index(lseg, j);
535         ds = nfs4_fl_prepare_ds(lseg, idx);
536         if (!ds)
537                 return PNFS_NOT_ATTEMPTED;
538
539         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
540         if (IS_ERR(ds_clnt))
541                 return PNFS_NOT_ATTEMPTED;
542
543         dprintk("%s ino %lu sync %d req %zu@%llu DS: %s cl_count %d\n",
544                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
545                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
546
547         hdr->pgio_done_cb = filelayout_write_done_cb;
548         atomic_inc(&ds->ds_clp->cl_count);
549         hdr->ds_clp = ds->ds_clp;
550         hdr->ds_commit_idx = idx;
551         fh = nfs4_fl_select_ds_fh(lseg, j);
552         if (fh)
553                 hdr->args.fh = fh;
554         hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
555
556         /* Perform an asynchronous write */
557         nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
558                           NFS_PROTO(hdr->inode), &filelayout_write_call_ops,
559                           sync, RPC_TASK_SOFTCONN);
560         return PNFS_ATTEMPTED;
561 }
562
563 /*
564  * filelayout_check_layout()
565  *
566  * Make sure layout segment parameters are sane WRT the device.
567  * At this point no generic layer initialization of the lseg has occurred,
568  * and nothing has been added to the layout_hdr cache.
569  *
570  */
571 static int
572 filelayout_check_layout(struct pnfs_layout_hdr *lo,
573                         struct nfs4_filelayout_segment *fl,
574                         struct nfs4_layoutget_res *lgr,
575                         struct nfs4_deviceid *id,
576                         gfp_t gfp_flags)
577 {
578         struct nfs4_deviceid_node *d;
579         struct nfs4_file_layout_dsaddr *dsaddr;
580         int status = -EINVAL;
581
582         dprintk("--> %s\n", __func__);
583
584         /* FIXME: remove this check when layout segment support is added */
585         if (lgr->range.offset != 0 ||
586             lgr->range.length != NFS4_MAX_UINT64) {
587                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
588                         __func__);
589                 goto out;
590         }
591
592         if (fl->pattern_offset > lgr->range.offset) {
593                 dprintk("%s pattern_offset %lld too large\n",
594                                 __func__, fl->pattern_offset);
595                 goto out;
596         }
597
598         if (!fl->stripe_unit) {
599                 dprintk("%s Invalid stripe unit (%u)\n",
600                         __func__, fl->stripe_unit);
601                 goto out;
602         }
603
604         /* find and reference the deviceid */
605         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), id,
606                         lo->plh_lc_cred, gfp_flags);
607         if (d == NULL)
608                 goto out;
609
610         dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
611         /* Found deviceid is unavailable */
612         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
613                 goto out_put;
614
615         fl->dsaddr = dsaddr;
616
617         if (fl->first_stripe_index >= dsaddr->stripe_count) {
618                 dprintk("%s Bad first_stripe_index %u\n",
619                                 __func__, fl->first_stripe_index);
620                 goto out_put;
621         }
622
623         if ((fl->stripe_type == STRIPE_SPARSE &&
624             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
625             (fl->stripe_type == STRIPE_DENSE &&
626             fl->num_fh != dsaddr->stripe_count)) {
627                 dprintk("%s num_fh %u not valid for given packing\n",
628                         __func__, fl->num_fh);
629                 goto out_put;
630         }
631
632         status = 0;
633 out:
634         dprintk("--> %s returns %d\n", __func__, status);
635         return status;
636 out_put:
637         nfs4_fl_put_deviceid(dsaddr);
638         goto out;
639 }
640
641 static void _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
642 {
643         int i;
644
645         if (fl->fh_array) {
646                 for (i = 0; i < fl->num_fh; i++) {
647                         if (!fl->fh_array[i])
648                                 break;
649                         kfree(fl->fh_array[i]);
650                 }
651                 kfree(fl->fh_array);
652         }
653         kfree(fl);
654 }
655
656 static int
657 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
658                          struct nfs4_filelayout_segment *fl,
659                          struct nfs4_layoutget_res *lgr,
660                          struct nfs4_deviceid *id,
661                          gfp_t gfp_flags)
662 {
663         struct xdr_stream stream;
664         struct xdr_buf buf;
665         struct page *scratch;
666         __be32 *p;
667         uint32_t nfl_util;
668         int i;
669
670         dprintk("%s: set_layout_map Begin\n", __func__);
671
672         scratch = alloc_page(gfp_flags);
673         if (!scratch)
674                 return -ENOMEM;
675
676         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
677         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
678
679         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
680          * num_fh (4) */
681         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
682         if (unlikely(!p))
683                 goto out_err;
684
685         memcpy(id, p, sizeof(*id));
686         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
687         nfs4_print_deviceid(id);
688
689         nfl_util = be32_to_cpup(p++);
690         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
691                 fl->commit_through_mds = 1;
692         if (nfl_util & NFL4_UFLG_DENSE)
693                 fl->stripe_type = STRIPE_DENSE;
694         else
695                 fl->stripe_type = STRIPE_SPARSE;
696         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
697
698         fl->first_stripe_index = be32_to_cpup(p++);
699         p = xdr_decode_hyper(p, &fl->pattern_offset);
700         fl->num_fh = be32_to_cpup(p++);
701
702         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
703                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
704                 fl->pattern_offset);
705
706         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
707          * Futher checking is done in filelayout_check_layout */
708         if (fl->num_fh >
709             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
710                 goto out_err;
711
712         if (fl->num_fh > 0) {
713                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
714                                        gfp_flags);
715                 if (!fl->fh_array)
716                         goto out_err;
717         }
718
719         for (i = 0; i < fl->num_fh; i++) {
720                 /* Do we want to use a mempool here? */
721                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
722                 if (!fl->fh_array[i])
723                         goto out_err;
724
725                 p = xdr_inline_decode(&stream, 4);
726                 if (unlikely(!p))
727                         goto out_err;
728                 fl->fh_array[i]->size = be32_to_cpup(p++);
729                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
730                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
731                                i, fl->fh_array[i]->size);
732                         goto out_err;
733                 }
734
735                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
736                 if (unlikely(!p))
737                         goto out_err;
738                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
739                 dprintk("DEBUG: %s: fh len %d\n", __func__,
740                         fl->fh_array[i]->size);
741         }
742
743         __free_page(scratch);
744         return 0;
745
746 out_err:
747         __free_page(scratch);
748         return -EIO;
749 }
750
751 static void
752 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
753 {
754         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
755
756         dprintk("--> %s\n", __func__);
757         nfs4_fl_put_deviceid(fl->dsaddr);
758         /* This assumes a single RW lseg */
759         if (lseg->pls_range.iomode == IOMODE_RW) {
760                 struct nfs4_filelayout *flo;
761
762                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
763                 flo->commit_info.nbuckets = 0;
764                 kfree(flo->commit_info.buckets);
765                 flo->commit_info.buckets = NULL;
766         }
767         _filelayout_free_lseg(fl);
768 }
769
770 static int
771 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
772                              struct nfs_commit_info *cinfo,
773                              gfp_t gfp_flags)
774 {
775         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
776         struct pnfs_commit_bucket *buckets;
777         int size, i;
778
779         if (fl->commit_through_mds)
780                 return 0;
781
782         size = (fl->stripe_type == STRIPE_SPARSE) ?
783                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
784
785         if (cinfo->ds->nbuckets >= size) {
786                 /* This assumes there is only one IOMODE_RW lseg.  What
787                  * we really want to do is have a layout_hdr level
788                  * dictionary of <multipath_list4, fh> keys, each
789                  * associated with a struct list_head, populated by calls
790                  * to filelayout_write_pagelist().
791                  * */
792                 return 0;
793         }
794
795         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
796                           gfp_flags);
797         if (!buckets)
798                 return -ENOMEM;
799         for (i = 0; i < size; i++) {
800                 INIT_LIST_HEAD(&buckets[i].written);
801                 INIT_LIST_HEAD(&buckets[i].committing);
802                 /* mark direct verifier as unset */
803                 buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
804         }
805
806         spin_lock(&cinfo->inode->i_lock);
807         if (cinfo->ds->nbuckets >= size)
808                 goto out;
809         for (i = 0; i < cinfo->ds->nbuckets; i++) {
810                 list_splice(&cinfo->ds->buckets[i].written,
811                             &buckets[i].written);
812                 list_splice(&cinfo->ds->buckets[i].committing,
813                             &buckets[i].committing);
814                 buckets[i].direct_verf.committed =
815                         cinfo->ds->buckets[i].direct_verf.committed;
816                 buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
817                 buckets[i].clseg = cinfo->ds->buckets[i].clseg;
818         }
819         swap(cinfo->ds->buckets, buckets);
820         cinfo->ds->nbuckets = size;
821 out:
822         spin_unlock(&cinfo->inode->i_lock);
823         kfree(buckets);
824         return 0;
825 }
826
827 static struct pnfs_layout_segment *
828 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
829                       struct nfs4_layoutget_res *lgr,
830                       gfp_t gfp_flags)
831 {
832         struct nfs4_filelayout_segment *fl;
833         int rc;
834         struct nfs4_deviceid id;
835
836         dprintk("--> %s\n", __func__);
837         fl = kzalloc(sizeof(*fl), gfp_flags);
838         if (!fl)
839                 return NULL;
840
841         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
842         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
843                 _filelayout_free_lseg(fl);
844                 return NULL;
845         }
846         return &fl->generic_hdr;
847 }
848
849 /*
850  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
851  *
852  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
853  * of bytes (maximum @req->wb_bytes) that can be coalesced.
854  */
855 static size_t
856 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
857                    struct nfs_page *req)
858 {
859         unsigned int size;
860         u64 p_stripe, r_stripe;
861         u32 stripe_offset;
862         u64 segment_offset = pgio->pg_lseg->pls_range.offset;
863         u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
864
865         /* calls nfs_generic_pg_test */
866         size = pnfs_generic_pg_test(pgio, prev, req);
867         if (!size)
868                 return 0;
869
870         /* see if req and prev are in the same stripe */
871         if (prev) {
872                 p_stripe = (u64)req_offset(prev) - segment_offset;
873                 r_stripe = (u64)req_offset(req) - segment_offset;
874                 do_div(p_stripe, stripe_unit);
875                 do_div(r_stripe, stripe_unit);
876
877                 if (p_stripe != r_stripe)
878                         return 0;
879         }
880
881         /* calculate remaining bytes in the current stripe */
882         div_u64_rem((u64)req_offset(req) - segment_offset,
883                         stripe_unit,
884                         &stripe_offset);
885         WARN_ON_ONCE(stripe_offset > stripe_unit);
886         if (stripe_offset >= stripe_unit)
887                 return 0;
888         return min(stripe_unit - (unsigned int)stripe_offset, size);
889 }
890
891 static void
892 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
893                         struct nfs_page *req)
894 {
895         if (!pgio->pg_lseg) {
896                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
897                                            req->wb_context,
898                                            0,
899                                            NFS4_MAX_UINT64,
900                                            IOMODE_READ,
901                                            false,
902                                            GFP_KERNEL);
903                 if (IS_ERR(pgio->pg_lseg)) {
904                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
905                         pgio->pg_lseg = NULL;
906                         return;
907                 }
908         }
909         /* If no lseg, fall back to read through mds */
910         if (pgio->pg_lseg == NULL)
911                 nfs_pageio_reset_read_mds(pgio);
912 }
913
914 static void
915 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
916                          struct nfs_page *req)
917 {
918         struct nfs_commit_info cinfo;
919         int status;
920
921         if (!pgio->pg_lseg) {
922                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
923                                            req->wb_context,
924                                            0,
925                                            NFS4_MAX_UINT64,
926                                            IOMODE_RW,
927                                            false,
928                                            GFP_NOFS);
929                 if (IS_ERR(pgio->pg_lseg)) {
930                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
931                         pgio->pg_lseg = NULL;
932                         return;
933                 }
934         }
935
936         /* If no lseg, fall back to write through mds */
937         if (pgio->pg_lseg == NULL)
938                 goto out_mds;
939         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
940         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
941         if (status < 0) {
942                 pnfs_put_lseg(pgio->pg_lseg);
943                 pgio->pg_lseg = NULL;
944                 goto out_mds;
945         }
946         return;
947 out_mds:
948         nfs_pageio_reset_write_mds(pgio);
949 }
950
951 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
952         .pg_init = filelayout_pg_init_read,
953         .pg_test = filelayout_pg_test,
954         .pg_doio = pnfs_generic_pg_readpages,
955         .pg_cleanup = pnfs_generic_pg_cleanup,
956 };
957
958 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
959         .pg_init = filelayout_pg_init_write,
960         .pg_test = filelayout_pg_test,
961         .pg_doio = pnfs_generic_pg_writepages,
962         .pg_cleanup = pnfs_generic_pg_cleanup,
963 };
964
965 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
966 {
967         if (fl->stripe_type == STRIPE_SPARSE)
968                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
969         else
970                 return j;
971 }
972
973 static void
974 filelayout_mark_request_commit(struct nfs_page *req,
975                                struct pnfs_layout_segment *lseg,
976                                struct nfs_commit_info *cinfo,
977                                u32 ds_commit_idx)
978
979 {
980         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
981         u32 i, j;
982
983         if (fl->commit_through_mds) {
984                 nfs_request_add_commit_list(req, cinfo);
985         } else {
986                 /* Note that we are calling nfs4_fl_calc_j_index on each page
987                  * that ends up being committed to a data server.  An attractive
988                  * alternative is to add a field to nfs_write_data and nfs_page
989                  * to store the value calculated in filelayout_write_pagelist
990                  * and just use that here.
991                  */
992                 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
993                 i = select_bucket_index(fl, j);
994                 pnfs_layout_mark_request_commit(req, lseg, cinfo, i);
995         }
996 }
997
998 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
999 {
1000         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1001
1002         if (flseg->stripe_type == STRIPE_SPARSE)
1003                 return i;
1004         else
1005                 return nfs4_fl_calc_ds_index(lseg, i);
1006 }
1007
1008 static struct nfs_fh *
1009 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1010 {
1011         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1012
1013         if (flseg->stripe_type == STRIPE_SPARSE) {
1014                 if (flseg->num_fh == 1)
1015                         i = 0;
1016                 else if (flseg->num_fh == 0)
1017                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1018                         return NULL;
1019         }
1020         return flseg->fh_array[i];
1021 }
1022
1023 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1024 {
1025         struct pnfs_layout_segment *lseg = data->lseg;
1026         struct nfs4_pnfs_ds *ds;
1027         struct rpc_clnt *ds_clnt;
1028         u32 idx;
1029         struct nfs_fh *fh;
1030
1031         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1032         ds = nfs4_fl_prepare_ds(lseg, idx);
1033         if (!ds)
1034                 goto out_err;
1035
1036         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1037         if (IS_ERR(ds_clnt))
1038                 goto out_err;
1039
1040         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1041                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1042         data->commit_done_cb = filelayout_commit_done_cb;
1043         atomic_inc(&ds->ds_clp->cl_count);
1044         data->ds_clp = ds->ds_clp;
1045         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1046         if (fh)
1047                 data->args.fh = fh;
1048         return nfs_initiate_commit(ds_clnt, data, NFS_PROTO(data->inode),
1049                                    &filelayout_commit_call_ops, how,
1050                                    RPC_TASK_SOFTCONN);
1051 out_err:
1052         pnfs_generic_prepare_to_resend_writes(data);
1053         pnfs_generic_commit_release(data);
1054         return -EAGAIN;
1055 }
1056
1057 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1058  *                                 for @page
1059  * @cinfo - commit info for current inode
1060  * @page - page to search for matching head request
1061  *
1062  * Returns a the head request if one is found, otherwise returns NULL.
1063  */
1064 static struct nfs_page *
1065 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1066 {
1067         struct nfs_page *freq, *t;
1068         struct pnfs_commit_bucket *b;
1069         int i;
1070
1071         /* Linearly search the commit lists for each bucket until a matching
1072          * request is found */
1073         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1074                 list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1075                         if (freq->wb_page == page)
1076                                 return freq->wb_head;
1077                 }
1078                 list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1079                         if (freq->wb_page == page)
1080                                 return freq->wb_head;
1081                 }
1082         }
1083
1084         return NULL;
1085 }
1086
1087 static int
1088 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1089                            int how, struct nfs_commit_info *cinfo)
1090 {
1091         return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1092                                             filelayout_initiate_commit);
1093 }
1094
1095 static struct nfs4_deviceid_node *
1096 filelayout_alloc_deviceid_node(struct nfs_server *server,
1097                 struct pnfs_device *pdev, gfp_t gfp_flags)
1098 {
1099         struct nfs4_file_layout_dsaddr *dsaddr;
1100
1101         dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1102         if (!dsaddr)
1103                 return NULL;
1104         return &dsaddr->id_node;
1105 }
1106
1107 static void
1108 filelayout_free_deviceid_node(struct nfs4_deviceid_node *d)
1109 {
1110         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1111 }
1112
1113 static struct pnfs_layout_hdr *
1114 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1115 {
1116         struct nfs4_filelayout *flo;
1117
1118         flo = kzalloc(sizeof(*flo), gfp_flags);
1119         return flo != NULL ? &flo->generic_hdr : NULL;
1120 }
1121
1122 static void
1123 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1124 {
1125         kfree(FILELAYOUT_FROM_HDR(lo));
1126 }
1127
1128 static struct pnfs_ds_commit_info *
1129 filelayout_get_ds_info(struct inode *inode)
1130 {
1131         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1132
1133         if (layout == NULL)
1134                 return NULL;
1135         else
1136                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1137 }
1138
1139 static struct pnfs_layoutdriver_type filelayout_type = {
1140         .id                     = LAYOUT_NFSV4_1_FILES,
1141         .name                   = "LAYOUT_NFSV4_1_FILES",
1142         .owner                  = THIS_MODULE,
1143         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1144         .free_layout_hdr        = filelayout_free_layout_hdr,
1145         .alloc_lseg             = filelayout_alloc_lseg,
1146         .free_lseg              = filelayout_free_lseg,
1147         .pg_read_ops            = &filelayout_pg_read_ops,
1148         .pg_write_ops           = &filelayout_pg_write_ops,
1149         .get_ds_info            = &filelayout_get_ds_info,
1150         .mark_request_commit    = filelayout_mark_request_commit,
1151         .clear_request_commit   = pnfs_generic_clear_request_commit,
1152         .scan_commit_lists      = pnfs_generic_scan_commit_lists,
1153         .recover_commit_reqs    = pnfs_generic_recover_commit_reqs,
1154         .search_commit_reqs     = filelayout_search_commit_reqs,
1155         .commit_pagelist        = filelayout_commit_pagelist,
1156         .read_pagelist          = filelayout_read_pagelist,
1157         .write_pagelist         = filelayout_write_pagelist,
1158         .alloc_deviceid_node    = filelayout_alloc_deviceid_node,
1159         .free_deviceid_node     = filelayout_free_deviceid_node,
1160         .sync                   = pnfs_nfs_generic_sync,
1161 };
1162
1163 static int __init nfs4filelayout_init(void)
1164 {
1165         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1166                __func__);
1167         return pnfs_register_layoutdriver(&filelayout_type);
1168 }
1169
1170 static void __exit nfs4filelayout_exit(void)
1171 {
1172         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1173                __func__);
1174         pnfs_unregister_layoutdriver(&filelayout_type);
1175 }
1176
1177 MODULE_ALIAS("nfs-layouttype4-1");
1178
1179 module_init(nfs4filelayout_init);
1180 module_exit(nfs4filelayout_exit);