xref: /linux/fs/nfs/filelayout/filelayout.c (revision 827634added7f38b7d724cab1dccdb2b004c13c3)
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 
36 #include <linux/sunrpc/metrics.h>
37 
38 #include "../nfs4session.h"
39 #include "../internal.h"
40 #include "../delegation.h"
41 #include "filelayout.h"
42 #include "../nfs4trace.h"
43 
44 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
45 
46 MODULE_LICENSE("GPL");
47 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
48 MODULE_DESCRIPTION("The NFSv4 file layout driver");
49 
50 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
51 
52 static loff_t
53 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
54 			    loff_t offset)
55 {
56 	u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
57 	u64 stripe_no;
58 	u32 rem;
59 
60 	offset -= flseg->pattern_offset;
61 	stripe_no = div_u64(offset, stripe_width);
62 	div_u64_rem(offset, flseg->stripe_unit, &rem);
63 
64 	return stripe_no * flseg->stripe_unit + rem;
65 }
66 
67 /* This function is used by the layout driver to calculate the
68  * offset of the file on the dserver based on whether the
69  * layout type is STRIPE_DENSE or STRIPE_SPARSE
70  */
71 static loff_t
72 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
73 {
74 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
75 
76 	switch (flseg->stripe_type) {
77 	case STRIPE_SPARSE:
78 		return offset;
79 
80 	case STRIPE_DENSE:
81 		return filelayout_get_dense_offset(flseg, offset);
82 	}
83 
84 	BUG();
85 }
86 
87 static void filelayout_reset_write(struct nfs_pgio_header *hdr)
88 {
89 	struct rpc_task *task = &hdr->task;
90 
91 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
92 		dprintk("%s Reset task %5u for i/o through MDS "
93 			"(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
94 			hdr->task.tk_pid,
95 			hdr->inode->i_sb->s_id,
96 			(unsigned long long)NFS_FILEID(hdr->inode),
97 			hdr->args.count,
98 			(unsigned long long)hdr->args.offset);
99 
100 		task->tk_status = pnfs_write_done_resend_to_mds(hdr);
101 	}
102 }
103 
104 static void filelayout_reset_read(struct nfs_pgio_header *hdr)
105 {
106 	struct rpc_task *task = &hdr->task;
107 
108 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
109 		dprintk("%s Reset task %5u for i/o through MDS "
110 			"(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
111 			hdr->task.tk_pid,
112 			hdr->inode->i_sb->s_id,
113 			(unsigned long long)NFS_FILEID(hdr->inode),
114 			hdr->args.count,
115 			(unsigned long long)hdr->args.offset);
116 
117 		task->tk_status = pnfs_read_done_resend_to_mds(hdr);
118 	}
119 }
120 
121 static int filelayout_async_handle_error(struct rpc_task *task,
122 					 struct nfs4_state *state,
123 					 struct nfs_client *clp,
124 					 struct pnfs_layout_segment *lseg)
125 {
126 	struct pnfs_layout_hdr *lo = lseg->pls_layout;
127 	struct inode *inode = lo->plh_inode;
128 	struct nfs_server *mds_server = NFS_SERVER(inode);
129 	struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
130 	struct nfs_client *mds_client = mds_server->nfs_client;
131 	struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
132 
133 	if (task->tk_status >= 0)
134 		return 0;
135 
136 	switch (task->tk_status) {
137 	/* MDS state errors */
138 	case -NFS4ERR_DELEG_REVOKED:
139 	case -NFS4ERR_ADMIN_REVOKED:
140 	case -NFS4ERR_BAD_STATEID:
141 	case -NFS4ERR_OPENMODE:
142 		if (state == NULL)
143 			break;
144 		if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
145 			goto out_bad_stateid;
146 		goto wait_on_recovery;
147 	case -NFS4ERR_EXPIRED:
148 		if (state != NULL) {
149 			if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
150 				goto out_bad_stateid;
151 		}
152 		nfs4_schedule_lease_recovery(mds_client);
153 		goto wait_on_recovery;
154 	/* DS session errors */
155 	case -NFS4ERR_BADSESSION:
156 	case -NFS4ERR_BADSLOT:
157 	case -NFS4ERR_BAD_HIGH_SLOT:
158 	case -NFS4ERR_DEADSESSION:
159 	case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
160 	case -NFS4ERR_SEQ_FALSE_RETRY:
161 	case -NFS4ERR_SEQ_MISORDERED:
162 		dprintk("%s ERROR %d, Reset session. Exchangeid "
163 			"flags 0x%x\n", __func__, task->tk_status,
164 			clp->cl_exchange_flags);
165 		nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
166 		break;
167 	case -NFS4ERR_DELAY:
168 	case -NFS4ERR_GRACE:
169 		rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
170 		break;
171 	case -NFS4ERR_RETRY_UNCACHED_REP:
172 		break;
173 	/* Invalidate Layout errors */
174 	case -NFS4ERR_PNFS_NO_LAYOUT:
175 	case -ESTALE:           /* mapped NFS4ERR_STALE */
176 	case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
177 	case -EISDIR:           /* mapped NFS4ERR_ISDIR */
178 	case -NFS4ERR_FHEXPIRED:
179 	case -NFS4ERR_WRONG_TYPE:
180 		dprintk("%s Invalid layout error %d\n", __func__,
181 			task->tk_status);
182 		/*
183 		 * Destroy layout so new i/o will get a new layout.
184 		 * Layout will not be destroyed until all current lseg
185 		 * references are put. Mark layout as invalid to resend failed
186 		 * i/o and all i/o waiting on the slot table to the MDS until
187 		 * layout is destroyed and a new valid layout is obtained.
188 		 */
189 		pnfs_destroy_layout(NFS_I(inode));
190 		rpc_wake_up(&tbl->slot_tbl_waitq);
191 		goto reset;
192 	/* RPC connection errors */
193 	case -ECONNREFUSED:
194 	case -EHOSTDOWN:
195 	case -EHOSTUNREACH:
196 	case -ENETUNREACH:
197 	case -EIO:
198 	case -ETIMEDOUT:
199 	case -EPIPE:
200 		dprintk("%s DS connection error %d\n", __func__,
201 			task->tk_status);
202 		nfs4_mark_deviceid_unavailable(devid);
203 		pnfs_error_mark_layout_for_return(inode, lseg);
204 		rpc_wake_up(&tbl->slot_tbl_waitq);
205 		/* fall through */
206 	default:
207 reset:
208 		dprintk("%s Retry through MDS. Error %d\n", __func__,
209 			task->tk_status);
210 		return -NFS4ERR_RESET_TO_MDS;
211 	}
212 out:
213 	task->tk_status = 0;
214 	return -EAGAIN;
215 out_bad_stateid:
216 	task->tk_status = -EIO;
217 	return 0;
218 wait_on_recovery:
219 	rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
220 	if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
221 		rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
222 	goto out;
223 }
224 
225 /* NFS_PROTO call done callback routines */
226 
227 static int filelayout_read_done_cb(struct rpc_task *task,
228 				struct nfs_pgio_header *hdr)
229 {
230 	int err;
231 
232 	trace_nfs4_pnfs_read(hdr, task->tk_status);
233 	err = filelayout_async_handle_error(task, hdr->args.context->state,
234 					    hdr->ds_clp, hdr->lseg);
235 
236 	switch (err) {
237 	case -NFS4ERR_RESET_TO_MDS:
238 		filelayout_reset_read(hdr);
239 		return task->tk_status;
240 	case -EAGAIN:
241 		rpc_restart_call_prepare(task);
242 		return -EAGAIN;
243 	}
244 
245 	return 0;
246 }
247 
248 /*
249  * We reference the rpc_cred of the first WRITE that triggers the need for
250  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
251  * rfc5661 is not clear about which credential should be used.
252  */
253 static void
254 filelayout_set_layoutcommit(struct nfs_pgio_header *hdr)
255 {
256 
257 	if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
258 	    hdr->res.verf->committed != NFS_DATA_SYNC)
259 		return;
260 
261 	pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
262 			hdr->mds_offset + hdr->res.count);
263 	dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
264 		(unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
265 }
266 
267 bool
268 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
269 {
270 	return filelayout_test_devid_invalid(node) ||
271 		nfs4_test_deviceid_unavailable(node);
272 }
273 
274 static bool
275 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
276 {
277 	struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
278 
279 	return filelayout_test_devid_unavailable(node);
280 }
281 
282 /*
283  * Call ops for the async read/write cases
284  * In the case of dense layouts, the offset needs to be reset to its
285  * original value.
286  */
287 static void filelayout_read_prepare(struct rpc_task *task, void *data)
288 {
289 	struct nfs_pgio_header *hdr = data;
290 
291 	if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
292 		rpc_exit(task, -EIO);
293 		return;
294 	}
295 	if (filelayout_reset_to_mds(hdr->lseg)) {
296 		dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
297 		filelayout_reset_read(hdr);
298 		rpc_exit(task, 0);
299 		return;
300 	}
301 	hdr->pgio_done_cb = filelayout_read_done_cb;
302 
303 	if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
304 			&hdr->args.seq_args,
305 			&hdr->res.seq_res,
306 			task))
307 		return;
308 	if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
309 			hdr->args.lock_context, FMODE_READ) == -EIO)
310 		rpc_exit(task, -EIO); /* lost lock, terminate I/O */
311 }
312 
313 static void filelayout_read_call_done(struct rpc_task *task, void *data)
314 {
315 	struct nfs_pgio_header *hdr = data;
316 
317 	dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
318 
319 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
320 	    task->tk_status == 0) {
321 		nfs41_sequence_done(task, &hdr->res.seq_res);
322 		return;
323 	}
324 
325 	/* Note this may cause RPC to be resent */
326 	hdr->mds_ops->rpc_call_done(task, data);
327 }
328 
329 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
330 {
331 	struct nfs_pgio_header *hdr = data;
332 
333 	rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
334 }
335 
336 static int filelayout_write_done_cb(struct rpc_task *task,
337 				struct nfs_pgio_header *hdr)
338 {
339 	int err;
340 
341 	trace_nfs4_pnfs_write(hdr, task->tk_status);
342 	err = filelayout_async_handle_error(task, hdr->args.context->state,
343 					    hdr->ds_clp, hdr->lseg);
344 
345 	switch (err) {
346 	case -NFS4ERR_RESET_TO_MDS:
347 		filelayout_reset_write(hdr);
348 		return task->tk_status;
349 	case -EAGAIN:
350 		rpc_restart_call_prepare(task);
351 		return -EAGAIN;
352 	}
353 
354 	filelayout_set_layoutcommit(hdr);
355 	return 0;
356 }
357 
358 static int filelayout_commit_done_cb(struct rpc_task *task,
359 				     struct nfs_commit_data *data)
360 {
361 	int err;
362 
363 	trace_nfs4_pnfs_commit_ds(data, task->tk_status);
364 	err = filelayout_async_handle_error(task, NULL, data->ds_clp,
365 					    data->lseg);
366 
367 	switch (err) {
368 	case -NFS4ERR_RESET_TO_MDS:
369 		pnfs_generic_prepare_to_resend_writes(data);
370 		return -EAGAIN;
371 	case -EAGAIN:
372 		rpc_restart_call_prepare(task);
373 		return -EAGAIN;
374 	}
375 
376 	if (data->verf.committed == NFS_UNSTABLE)
377 		pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
378 
379 	return 0;
380 }
381 
382 static void filelayout_write_prepare(struct rpc_task *task, void *data)
383 {
384 	struct nfs_pgio_header *hdr = data;
385 
386 	if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
387 		rpc_exit(task, -EIO);
388 		return;
389 	}
390 	if (filelayout_reset_to_mds(hdr->lseg)) {
391 		dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
392 		filelayout_reset_write(hdr);
393 		rpc_exit(task, 0);
394 		return;
395 	}
396 	if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
397 			&hdr->args.seq_args,
398 			&hdr->res.seq_res,
399 			task))
400 		return;
401 	if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
402 			hdr->args.lock_context, FMODE_WRITE) == -EIO)
403 		rpc_exit(task, -EIO); /* lost lock, terminate I/O */
404 }
405 
406 static void filelayout_write_call_done(struct rpc_task *task, void *data)
407 {
408 	struct nfs_pgio_header *hdr = data;
409 
410 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
411 	    task->tk_status == 0) {
412 		nfs41_sequence_done(task, &hdr->res.seq_res);
413 		return;
414 	}
415 
416 	/* Note this may cause RPC to be resent */
417 	hdr->mds_ops->rpc_call_done(task, data);
418 }
419 
420 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
421 {
422 	struct nfs_pgio_header *hdr = data;
423 
424 	rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
425 }
426 
427 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
428 {
429 	struct nfs_commit_data *wdata = data;
430 
431 	nfs41_setup_sequence(wdata->ds_clp->cl_session,
432 			&wdata->args.seq_args,
433 			&wdata->res.seq_res,
434 			task);
435 }
436 
437 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
438 {
439 	struct nfs_commit_data *cdata = data;
440 
441 	rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
442 }
443 
444 static const struct rpc_call_ops filelayout_read_call_ops = {
445 	.rpc_call_prepare = filelayout_read_prepare,
446 	.rpc_call_done = filelayout_read_call_done,
447 	.rpc_count_stats = filelayout_read_count_stats,
448 	.rpc_release = pnfs_generic_rw_release,
449 };
450 
451 static const struct rpc_call_ops filelayout_write_call_ops = {
452 	.rpc_call_prepare = filelayout_write_prepare,
453 	.rpc_call_done = filelayout_write_call_done,
454 	.rpc_count_stats = filelayout_write_count_stats,
455 	.rpc_release = pnfs_generic_rw_release,
456 };
457 
458 static const struct rpc_call_ops filelayout_commit_call_ops = {
459 	.rpc_call_prepare = filelayout_commit_prepare,
460 	.rpc_call_done = pnfs_generic_write_commit_done,
461 	.rpc_count_stats = filelayout_commit_count_stats,
462 	.rpc_release = pnfs_generic_commit_release,
463 };
464 
465 static enum pnfs_try_status
466 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
467 {
468 	struct pnfs_layout_segment *lseg = hdr->lseg;
469 	struct nfs4_pnfs_ds *ds;
470 	struct rpc_clnt *ds_clnt;
471 	loff_t offset = hdr->args.offset;
472 	u32 j, idx;
473 	struct nfs_fh *fh;
474 
475 	dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
476 		__func__, hdr->inode->i_ino,
477 		hdr->args.pgbase, (size_t)hdr->args.count, offset);
478 
479 	/* Retrieve the correct rpc_client for the byte range */
480 	j = nfs4_fl_calc_j_index(lseg, offset);
481 	idx = nfs4_fl_calc_ds_index(lseg, j);
482 	ds = nfs4_fl_prepare_ds(lseg, idx);
483 	if (!ds)
484 		return PNFS_NOT_ATTEMPTED;
485 
486 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
487 	if (IS_ERR(ds_clnt))
488 		return PNFS_NOT_ATTEMPTED;
489 
490 	dprintk("%s USE DS: %s cl_count %d\n", __func__,
491 		ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
492 
493 	/* No multipath support. Use first DS */
494 	atomic_inc(&ds->ds_clp->cl_count);
495 	hdr->ds_clp = ds->ds_clp;
496 	hdr->ds_commit_idx = idx;
497 	fh = nfs4_fl_select_ds_fh(lseg, j);
498 	if (fh)
499 		hdr->args.fh = fh;
500 
501 	hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
502 	hdr->mds_offset = offset;
503 
504 	/* Perform an asynchronous read to ds */
505 	nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
506 			  NFS_PROTO(hdr->inode), &filelayout_read_call_ops,
507 			  0, RPC_TASK_SOFTCONN);
508 	return PNFS_ATTEMPTED;
509 }
510 
511 /* Perform async writes. */
512 static enum pnfs_try_status
513 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
514 {
515 	struct pnfs_layout_segment *lseg = hdr->lseg;
516 	struct nfs4_pnfs_ds *ds;
517 	struct rpc_clnt *ds_clnt;
518 	loff_t offset = hdr->args.offset;
519 	u32 j, idx;
520 	struct nfs_fh *fh;
521 
522 	/* Retrieve the correct rpc_client for the byte range */
523 	j = nfs4_fl_calc_j_index(lseg, offset);
524 	idx = nfs4_fl_calc_ds_index(lseg, j);
525 	ds = nfs4_fl_prepare_ds(lseg, idx);
526 	if (!ds)
527 		return PNFS_NOT_ATTEMPTED;
528 
529 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
530 	if (IS_ERR(ds_clnt))
531 		return PNFS_NOT_ATTEMPTED;
532 
533 	dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
534 		__func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
535 		offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
536 
537 	hdr->pgio_done_cb = filelayout_write_done_cb;
538 	atomic_inc(&ds->ds_clp->cl_count);
539 	hdr->ds_clp = ds->ds_clp;
540 	hdr->ds_commit_idx = idx;
541 	fh = nfs4_fl_select_ds_fh(lseg, j);
542 	if (fh)
543 		hdr->args.fh = fh;
544 	hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
545 
546 	/* Perform an asynchronous write */
547 	nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
548 			  NFS_PROTO(hdr->inode), &filelayout_write_call_ops,
549 			  sync, RPC_TASK_SOFTCONN);
550 	return PNFS_ATTEMPTED;
551 }
552 
553 /*
554  * filelayout_check_layout()
555  *
556  * Make sure layout segment parameters are sane WRT the device.
557  * At this point no generic layer initialization of the lseg has occurred,
558  * and nothing has been added to the layout_hdr cache.
559  *
560  */
561 static int
562 filelayout_check_layout(struct pnfs_layout_hdr *lo,
563 			struct nfs4_filelayout_segment *fl,
564 			struct nfs4_layoutget_res *lgr,
565 			struct nfs4_deviceid *id,
566 			gfp_t gfp_flags)
567 {
568 	struct nfs4_deviceid_node *d;
569 	struct nfs4_file_layout_dsaddr *dsaddr;
570 	int status = -EINVAL;
571 
572 	dprintk("--> %s\n", __func__);
573 
574 	/* FIXME: remove this check when layout segment support is added */
575 	if (lgr->range.offset != 0 ||
576 	    lgr->range.length != NFS4_MAX_UINT64) {
577 		dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
578 			__func__);
579 		goto out;
580 	}
581 
582 	if (fl->pattern_offset > lgr->range.offset) {
583 		dprintk("%s pattern_offset %lld too large\n",
584 				__func__, fl->pattern_offset);
585 		goto out;
586 	}
587 
588 	if (!fl->stripe_unit) {
589 		dprintk("%s Invalid stripe unit (%u)\n",
590 			__func__, fl->stripe_unit);
591 		goto out;
592 	}
593 
594 	/* find and reference the deviceid */
595 	d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), id,
596 			lo->plh_lc_cred, gfp_flags);
597 	if (d == NULL)
598 		goto out;
599 
600 	dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
601 	/* Found deviceid is unavailable */
602 	if (filelayout_test_devid_unavailable(&dsaddr->id_node))
603 		goto out_put;
604 
605 	fl->dsaddr = dsaddr;
606 
607 	if (fl->first_stripe_index >= dsaddr->stripe_count) {
608 		dprintk("%s Bad first_stripe_index %u\n",
609 				__func__, fl->first_stripe_index);
610 		goto out_put;
611 	}
612 
613 	if ((fl->stripe_type == STRIPE_SPARSE &&
614 	    fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
615 	    (fl->stripe_type == STRIPE_DENSE &&
616 	    fl->num_fh != dsaddr->stripe_count)) {
617 		dprintk("%s num_fh %u not valid for given packing\n",
618 			__func__, fl->num_fh);
619 		goto out_put;
620 	}
621 
622 	status = 0;
623 out:
624 	dprintk("--> %s returns %d\n", __func__, status);
625 	return status;
626 out_put:
627 	nfs4_fl_put_deviceid(dsaddr);
628 	goto out;
629 }
630 
631 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
632 {
633 	int i;
634 
635 	for (i = 0; i < fl->num_fh; i++) {
636 		if (!fl->fh_array[i])
637 			break;
638 		kfree(fl->fh_array[i]);
639 	}
640 	kfree(fl->fh_array);
641 	fl->fh_array = NULL;
642 }
643 
644 static void
645 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
646 {
647 	filelayout_free_fh_array(fl);
648 	kfree(fl);
649 }
650 
651 static int
652 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
653 			 struct nfs4_filelayout_segment *fl,
654 			 struct nfs4_layoutget_res *lgr,
655 			 struct nfs4_deviceid *id,
656 			 gfp_t gfp_flags)
657 {
658 	struct xdr_stream stream;
659 	struct xdr_buf buf;
660 	struct page *scratch;
661 	__be32 *p;
662 	uint32_t nfl_util;
663 	int i;
664 
665 	dprintk("%s: set_layout_map Begin\n", __func__);
666 
667 	scratch = alloc_page(gfp_flags);
668 	if (!scratch)
669 		return -ENOMEM;
670 
671 	xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
672 	xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
673 
674 	/* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
675 	 * num_fh (4) */
676 	p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
677 	if (unlikely(!p))
678 		goto out_err;
679 
680 	memcpy(id, p, sizeof(*id));
681 	p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
682 	nfs4_print_deviceid(id);
683 
684 	nfl_util = be32_to_cpup(p++);
685 	if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
686 		fl->commit_through_mds = 1;
687 	if (nfl_util & NFL4_UFLG_DENSE)
688 		fl->stripe_type = STRIPE_DENSE;
689 	else
690 		fl->stripe_type = STRIPE_SPARSE;
691 	fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
692 
693 	fl->first_stripe_index = be32_to_cpup(p++);
694 	p = xdr_decode_hyper(p, &fl->pattern_offset);
695 	fl->num_fh = be32_to_cpup(p++);
696 
697 	dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
698 		__func__, nfl_util, fl->num_fh, fl->first_stripe_index,
699 		fl->pattern_offset);
700 
701 	/* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
702 	 * Futher checking is done in filelayout_check_layout */
703 	if (fl->num_fh >
704 	    max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
705 		goto out_err;
706 
707 	if (fl->num_fh > 0) {
708 		fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
709 				       gfp_flags);
710 		if (!fl->fh_array)
711 			goto out_err;
712 	}
713 
714 	for (i = 0; i < fl->num_fh; i++) {
715 		/* Do we want to use a mempool here? */
716 		fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
717 		if (!fl->fh_array[i])
718 			goto out_err_free;
719 
720 		p = xdr_inline_decode(&stream, 4);
721 		if (unlikely(!p))
722 			goto out_err_free;
723 		fl->fh_array[i]->size = be32_to_cpup(p++);
724 		if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
725 			printk(KERN_ERR "NFS: Too big fh %d received %d\n",
726 			       i, fl->fh_array[i]->size);
727 			goto out_err_free;
728 		}
729 
730 		p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
731 		if (unlikely(!p))
732 			goto out_err_free;
733 		memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
734 		dprintk("DEBUG: %s: fh len %d\n", __func__,
735 			fl->fh_array[i]->size);
736 	}
737 
738 	__free_page(scratch);
739 	return 0;
740 
741 out_err_free:
742 	filelayout_free_fh_array(fl);
743 out_err:
744 	__free_page(scratch);
745 	return -EIO;
746 }
747 
748 static void
749 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
750 {
751 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
752 
753 	dprintk("--> %s\n", __func__);
754 	nfs4_fl_put_deviceid(fl->dsaddr);
755 	/* This assumes a single RW lseg */
756 	if (lseg->pls_range.iomode == IOMODE_RW) {
757 		struct nfs4_filelayout *flo;
758 
759 		flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
760 		flo->commit_info.nbuckets = 0;
761 		kfree(flo->commit_info.buckets);
762 		flo->commit_info.buckets = NULL;
763 	}
764 	_filelayout_free_lseg(fl);
765 }
766 
767 static int
768 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
769 			     struct nfs_commit_info *cinfo,
770 			     gfp_t gfp_flags)
771 {
772 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
773 	struct pnfs_commit_bucket *buckets;
774 	int size, i;
775 
776 	if (fl->commit_through_mds)
777 		return 0;
778 
779 	size = (fl->stripe_type == STRIPE_SPARSE) ?
780 		fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
781 
782 	if (cinfo->ds->nbuckets >= size) {
783 		/* This assumes there is only one IOMODE_RW lseg.  What
784 		 * we really want to do is have a layout_hdr level
785 		 * dictionary of <multipath_list4, fh> keys, each
786 		 * associated with a struct list_head, populated by calls
787 		 * to filelayout_write_pagelist().
788 		 * */
789 		return 0;
790 	}
791 
792 	buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
793 			  gfp_flags);
794 	if (!buckets)
795 		return -ENOMEM;
796 	for (i = 0; i < size; i++) {
797 		INIT_LIST_HEAD(&buckets[i].written);
798 		INIT_LIST_HEAD(&buckets[i].committing);
799 		/* mark direct verifier as unset */
800 		buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
801 	}
802 
803 	spin_lock(cinfo->lock);
804 	if (cinfo->ds->nbuckets >= size)
805 		goto out;
806 	for (i = 0; i < cinfo->ds->nbuckets; i++) {
807 		list_splice(&cinfo->ds->buckets[i].written,
808 			    &buckets[i].written);
809 		list_splice(&cinfo->ds->buckets[i].committing,
810 			    &buckets[i].committing);
811 		buckets[i].direct_verf.committed =
812 			cinfo->ds->buckets[i].direct_verf.committed;
813 		buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
814 		buckets[i].clseg = cinfo->ds->buckets[i].clseg;
815 	}
816 	swap(cinfo->ds->buckets, buckets);
817 	cinfo->ds->nbuckets = size;
818 out:
819 	spin_unlock(cinfo->lock);
820 	kfree(buckets);
821 	return 0;
822 }
823 
824 static struct pnfs_layout_segment *
825 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
826 		      struct nfs4_layoutget_res *lgr,
827 		      gfp_t gfp_flags)
828 {
829 	struct nfs4_filelayout_segment *fl;
830 	int rc;
831 	struct nfs4_deviceid id;
832 
833 	dprintk("--> %s\n", __func__);
834 	fl = kzalloc(sizeof(*fl), gfp_flags);
835 	if (!fl)
836 		return NULL;
837 
838 	rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
839 	if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
840 		_filelayout_free_lseg(fl);
841 		return NULL;
842 	}
843 	return &fl->generic_hdr;
844 }
845 
846 /*
847  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
848  *
849  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
850  * of bytes (maximum @req->wb_bytes) that can be coalesced.
851  */
852 static size_t
853 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
854 		   struct nfs_page *req)
855 {
856 	unsigned int size;
857 	u64 p_stripe, r_stripe;
858 	u32 stripe_offset;
859 	u64 segment_offset = pgio->pg_lseg->pls_range.offset;
860 	u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
861 
862 	/* calls nfs_generic_pg_test */
863 	size = pnfs_generic_pg_test(pgio, prev, req);
864 	if (!size)
865 		return 0;
866 
867 	/* see if req and prev are in the same stripe */
868 	if (prev) {
869 		p_stripe = (u64)req_offset(prev) - segment_offset;
870 		r_stripe = (u64)req_offset(req) - segment_offset;
871 		do_div(p_stripe, stripe_unit);
872 		do_div(r_stripe, stripe_unit);
873 
874 		if (p_stripe != r_stripe)
875 			return 0;
876 	}
877 
878 	/* calculate remaining bytes in the current stripe */
879 	div_u64_rem((u64)req_offset(req) - segment_offset,
880 			stripe_unit,
881 			&stripe_offset);
882 	WARN_ON_ONCE(stripe_offset > stripe_unit);
883 	if (stripe_offset >= stripe_unit)
884 		return 0;
885 	return min(stripe_unit - (unsigned int)stripe_offset, size);
886 }
887 
888 static void
889 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
890 			struct nfs_page *req)
891 {
892 	if (!pgio->pg_lseg)
893 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
894 					   req->wb_context,
895 					   0,
896 					   NFS4_MAX_UINT64,
897 					   IOMODE_READ,
898 					   GFP_KERNEL);
899 	/* If no lseg, fall back to read through mds */
900 	if (pgio->pg_lseg == NULL)
901 		nfs_pageio_reset_read_mds(pgio);
902 }
903 
904 static void
905 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
906 			 struct nfs_page *req)
907 {
908 	struct nfs_commit_info cinfo;
909 	int status;
910 
911 	if (!pgio->pg_lseg)
912 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
913 					   req->wb_context,
914 					   0,
915 					   NFS4_MAX_UINT64,
916 					   IOMODE_RW,
917 					   GFP_NOFS);
918 	/* If no lseg, fall back to write through mds */
919 	if (pgio->pg_lseg == NULL)
920 		goto out_mds;
921 	nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
922 	status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
923 	if (status < 0) {
924 		pnfs_put_lseg(pgio->pg_lseg);
925 		pgio->pg_lseg = NULL;
926 		goto out_mds;
927 	}
928 	return;
929 out_mds:
930 	nfs_pageio_reset_write_mds(pgio);
931 }
932 
933 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
934 	.pg_init = filelayout_pg_init_read,
935 	.pg_test = filelayout_pg_test,
936 	.pg_doio = pnfs_generic_pg_readpages,
937 	.pg_cleanup = pnfs_generic_pg_cleanup,
938 };
939 
940 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
941 	.pg_init = filelayout_pg_init_write,
942 	.pg_test = filelayout_pg_test,
943 	.pg_doio = pnfs_generic_pg_writepages,
944 	.pg_cleanup = pnfs_generic_pg_cleanup,
945 };
946 
947 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
948 {
949 	if (fl->stripe_type == STRIPE_SPARSE)
950 		return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
951 	else
952 		return j;
953 }
954 
955 static void
956 filelayout_mark_request_commit(struct nfs_page *req,
957 			       struct pnfs_layout_segment *lseg,
958 			       struct nfs_commit_info *cinfo,
959 			       u32 ds_commit_idx)
960 
961 {
962 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
963 	u32 i, j;
964 
965 	if (fl->commit_through_mds) {
966 		nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
967 	} else {
968 		/* Note that we are calling nfs4_fl_calc_j_index on each page
969 		 * that ends up being committed to a data server.  An attractive
970 		 * alternative is to add a field to nfs_write_data and nfs_page
971 		 * to store the value calculated in filelayout_write_pagelist
972 		 * and just use that here.
973 		 */
974 		j = nfs4_fl_calc_j_index(lseg, req_offset(req));
975 		i = select_bucket_index(fl, j);
976 		pnfs_layout_mark_request_commit(req, lseg, cinfo, i);
977 	}
978 }
979 
980 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
981 {
982 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
983 
984 	if (flseg->stripe_type == STRIPE_SPARSE)
985 		return i;
986 	else
987 		return nfs4_fl_calc_ds_index(lseg, i);
988 }
989 
990 static struct nfs_fh *
991 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
992 {
993 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
994 
995 	if (flseg->stripe_type == STRIPE_SPARSE) {
996 		if (flseg->num_fh == 1)
997 			i = 0;
998 		else if (flseg->num_fh == 0)
999 			/* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1000 			return NULL;
1001 	}
1002 	return flseg->fh_array[i];
1003 }
1004 
1005 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1006 {
1007 	struct pnfs_layout_segment *lseg = data->lseg;
1008 	struct nfs4_pnfs_ds *ds;
1009 	struct rpc_clnt *ds_clnt;
1010 	u32 idx;
1011 	struct nfs_fh *fh;
1012 
1013 	idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1014 	ds = nfs4_fl_prepare_ds(lseg, idx);
1015 	if (!ds)
1016 		goto out_err;
1017 
1018 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1019 	if (IS_ERR(ds_clnt))
1020 		goto out_err;
1021 
1022 	dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1023 		data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1024 	data->commit_done_cb = filelayout_commit_done_cb;
1025 	atomic_inc(&ds->ds_clp->cl_count);
1026 	data->ds_clp = ds->ds_clp;
1027 	fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1028 	if (fh)
1029 		data->args.fh = fh;
1030 	return nfs_initiate_commit(ds_clnt, data, NFS_PROTO(data->inode),
1031 				   &filelayout_commit_call_ops, how,
1032 				   RPC_TASK_SOFTCONN);
1033 out_err:
1034 	pnfs_generic_prepare_to_resend_writes(data);
1035 	pnfs_generic_commit_release(data);
1036 	return -EAGAIN;
1037 }
1038 
1039 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1040  *				   for @page
1041  * @cinfo - commit info for current inode
1042  * @page - page to search for matching head request
1043  *
1044  * Returns a the head request if one is found, otherwise returns NULL.
1045  */
1046 static struct nfs_page *
1047 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1048 {
1049 	struct nfs_page *freq, *t;
1050 	struct pnfs_commit_bucket *b;
1051 	int i;
1052 
1053 	/* Linearly search the commit lists for each bucket until a matching
1054 	 * request is found */
1055 	for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1056 		list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1057 			if (freq->wb_page == page)
1058 				return freq->wb_head;
1059 		}
1060 		list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1061 			if (freq->wb_page == page)
1062 				return freq->wb_head;
1063 		}
1064 	}
1065 
1066 	return NULL;
1067 }
1068 
1069 static int
1070 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1071 			   int how, struct nfs_commit_info *cinfo)
1072 {
1073 	return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1074 					    filelayout_initiate_commit);
1075 }
1076 
1077 static struct nfs4_deviceid_node *
1078 filelayout_alloc_deviceid_node(struct nfs_server *server,
1079 		struct pnfs_device *pdev, gfp_t gfp_flags)
1080 {
1081 	struct nfs4_file_layout_dsaddr *dsaddr;
1082 
1083 	dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1084 	if (!dsaddr)
1085 		return NULL;
1086 	return &dsaddr->id_node;
1087 }
1088 
1089 static void
1090 filelayout_free_deviceid_node(struct nfs4_deviceid_node *d)
1091 {
1092 	nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1093 }
1094 
1095 static struct pnfs_layout_hdr *
1096 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1097 {
1098 	struct nfs4_filelayout *flo;
1099 
1100 	flo = kzalloc(sizeof(*flo), gfp_flags);
1101 	return flo != NULL ? &flo->generic_hdr : NULL;
1102 }
1103 
1104 static void
1105 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1106 {
1107 	kfree(FILELAYOUT_FROM_HDR(lo));
1108 }
1109 
1110 static struct pnfs_ds_commit_info *
1111 filelayout_get_ds_info(struct inode *inode)
1112 {
1113 	struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1114 
1115 	if (layout == NULL)
1116 		return NULL;
1117 	else
1118 		return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1119 }
1120 
1121 static struct pnfs_layoutdriver_type filelayout_type = {
1122 	.id			= LAYOUT_NFSV4_1_FILES,
1123 	.name			= "LAYOUT_NFSV4_1_FILES",
1124 	.owner			= THIS_MODULE,
1125 	.alloc_layout_hdr	= filelayout_alloc_layout_hdr,
1126 	.free_layout_hdr	= filelayout_free_layout_hdr,
1127 	.alloc_lseg		= filelayout_alloc_lseg,
1128 	.free_lseg		= filelayout_free_lseg,
1129 	.pg_read_ops		= &filelayout_pg_read_ops,
1130 	.pg_write_ops		= &filelayout_pg_write_ops,
1131 	.get_ds_info		= &filelayout_get_ds_info,
1132 	.mark_request_commit	= filelayout_mark_request_commit,
1133 	.clear_request_commit	= pnfs_generic_clear_request_commit,
1134 	.scan_commit_lists	= pnfs_generic_scan_commit_lists,
1135 	.recover_commit_reqs	= pnfs_generic_recover_commit_reqs,
1136 	.search_commit_reqs	= filelayout_search_commit_reqs,
1137 	.commit_pagelist	= filelayout_commit_pagelist,
1138 	.read_pagelist		= filelayout_read_pagelist,
1139 	.write_pagelist		= filelayout_write_pagelist,
1140 	.alloc_deviceid_node	= filelayout_alloc_deviceid_node,
1141 	.free_deviceid_node	= filelayout_free_deviceid_node,
1142 	.sync			= pnfs_nfs_generic_sync,
1143 };
1144 
1145 static int __init nfs4filelayout_init(void)
1146 {
1147 	printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1148 	       __func__);
1149 	return pnfs_register_layoutdriver(&filelayout_type);
1150 }
1151 
1152 static void __exit nfs4filelayout_exit(void)
1153 {
1154 	printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1155 	       __func__);
1156 	pnfs_unregister_layoutdriver(&filelayout_type);
1157 }
1158 
1159 MODULE_ALIAS("nfs-layouttype4-1");
1160 
1161 module_init(nfs4filelayout_init);
1162 module_exit(nfs4filelayout_exit);
1163