xref: /linux/fs/nfs/filelayout/filelayout.c (revision 2ba9268dd603d23e17643437b2246acb6844953b)
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);
262 	dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
263 		(unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
264 }
265 
266 bool
267 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
268 {
269 	return filelayout_test_devid_invalid(node) ||
270 		nfs4_test_deviceid_unavailable(node);
271 }
272 
273 static bool
274 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
275 {
276 	struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
277 
278 	return filelayout_test_devid_unavailable(node);
279 }
280 
281 /*
282  * Call ops for the async read/write cases
283  * In the case of dense layouts, the offset needs to be reset to its
284  * original value.
285  */
286 static void filelayout_read_prepare(struct rpc_task *task, void *data)
287 {
288 	struct nfs_pgio_header *hdr = data;
289 
290 	if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
291 		rpc_exit(task, -EIO);
292 		return;
293 	}
294 	if (filelayout_reset_to_mds(hdr->lseg)) {
295 		dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
296 		filelayout_reset_read(hdr);
297 		rpc_exit(task, 0);
298 		return;
299 	}
300 	hdr->pgio_done_cb = filelayout_read_done_cb;
301 
302 	if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
303 			&hdr->args.seq_args,
304 			&hdr->res.seq_res,
305 			task))
306 		return;
307 	if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
308 			hdr->args.lock_context, FMODE_READ) == -EIO)
309 		rpc_exit(task, -EIO); /* lost lock, terminate I/O */
310 }
311 
312 static void filelayout_read_call_done(struct rpc_task *task, void *data)
313 {
314 	struct nfs_pgio_header *hdr = data;
315 
316 	dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
317 
318 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
319 	    task->tk_status == 0) {
320 		nfs41_sequence_done(task, &hdr->res.seq_res);
321 		return;
322 	}
323 
324 	/* Note this may cause RPC to be resent */
325 	hdr->mds_ops->rpc_call_done(task, data);
326 }
327 
328 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
329 {
330 	struct nfs_pgio_header *hdr = data;
331 
332 	rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
333 }
334 
335 static int filelayout_write_done_cb(struct rpc_task *task,
336 				struct nfs_pgio_header *hdr)
337 {
338 	int err;
339 
340 	trace_nfs4_pnfs_write(hdr, task->tk_status);
341 	err = filelayout_async_handle_error(task, hdr->args.context->state,
342 					    hdr->ds_clp, hdr->lseg);
343 
344 	switch (err) {
345 	case -NFS4ERR_RESET_TO_MDS:
346 		filelayout_reset_write(hdr);
347 		return task->tk_status;
348 	case -EAGAIN:
349 		rpc_restart_call_prepare(task);
350 		return -EAGAIN;
351 	}
352 
353 	filelayout_set_layoutcommit(hdr);
354 	return 0;
355 }
356 
357 static int filelayout_commit_done_cb(struct rpc_task *task,
358 				     struct nfs_commit_data *data)
359 {
360 	int err;
361 
362 	trace_nfs4_pnfs_commit_ds(data, task->tk_status);
363 	err = filelayout_async_handle_error(task, NULL, data->ds_clp,
364 					    data->lseg);
365 
366 	switch (err) {
367 	case -NFS4ERR_RESET_TO_MDS:
368 		pnfs_generic_prepare_to_resend_writes(data);
369 		return -EAGAIN;
370 	case -EAGAIN:
371 		rpc_restart_call_prepare(task);
372 		return -EAGAIN;
373 	}
374 
375 	if (data->verf.committed == NFS_UNSTABLE)
376 		pnfs_commit_set_layoutcommit(data);
377 
378 	return 0;
379 }
380 
381 static void filelayout_write_prepare(struct rpc_task *task, void *data)
382 {
383 	struct nfs_pgio_header *hdr = data;
384 
385 	if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
386 		rpc_exit(task, -EIO);
387 		return;
388 	}
389 	if (filelayout_reset_to_mds(hdr->lseg)) {
390 		dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
391 		filelayout_reset_write(hdr);
392 		rpc_exit(task, 0);
393 		return;
394 	}
395 	if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
396 			&hdr->args.seq_args,
397 			&hdr->res.seq_res,
398 			task))
399 		return;
400 	if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
401 			hdr->args.lock_context, FMODE_WRITE) == -EIO)
402 		rpc_exit(task, -EIO); /* lost lock, terminate I/O */
403 }
404 
405 static void filelayout_write_call_done(struct rpc_task *task, void *data)
406 {
407 	struct nfs_pgio_header *hdr = data;
408 
409 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
410 	    task->tk_status == 0) {
411 		nfs41_sequence_done(task, &hdr->res.seq_res);
412 		return;
413 	}
414 
415 	/* Note this may cause RPC to be resent */
416 	hdr->mds_ops->rpc_call_done(task, data);
417 }
418 
419 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
420 {
421 	struct nfs_pgio_header *hdr = data;
422 
423 	rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
424 }
425 
426 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
427 {
428 	struct nfs_commit_data *wdata = data;
429 
430 	nfs41_setup_sequence(wdata->ds_clp->cl_session,
431 			&wdata->args.seq_args,
432 			&wdata->res.seq_res,
433 			task);
434 }
435 
436 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
437 {
438 	struct nfs_commit_data *cdata = data;
439 
440 	rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
441 }
442 
443 static const struct rpc_call_ops filelayout_read_call_ops = {
444 	.rpc_call_prepare = filelayout_read_prepare,
445 	.rpc_call_done = filelayout_read_call_done,
446 	.rpc_count_stats = filelayout_read_count_stats,
447 	.rpc_release = pnfs_generic_rw_release,
448 };
449 
450 static const struct rpc_call_ops filelayout_write_call_ops = {
451 	.rpc_call_prepare = filelayout_write_prepare,
452 	.rpc_call_done = filelayout_write_call_done,
453 	.rpc_count_stats = filelayout_write_count_stats,
454 	.rpc_release = pnfs_generic_rw_release,
455 };
456 
457 static const struct rpc_call_ops filelayout_commit_call_ops = {
458 	.rpc_call_prepare = filelayout_commit_prepare,
459 	.rpc_call_done = pnfs_generic_write_commit_done,
460 	.rpc_count_stats = filelayout_commit_count_stats,
461 	.rpc_release = pnfs_generic_commit_release,
462 };
463 
464 static enum pnfs_try_status
465 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
466 {
467 	struct pnfs_layout_segment *lseg = hdr->lseg;
468 	struct nfs4_pnfs_ds *ds;
469 	struct rpc_clnt *ds_clnt;
470 	loff_t offset = hdr->args.offset;
471 	u32 j, idx;
472 	struct nfs_fh *fh;
473 
474 	dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
475 		__func__, hdr->inode->i_ino,
476 		hdr->args.pgbase, (size_t)hdr->args.count, offset);
477 
478 	/* Retrieve the correct rpc_client for the byte range */
479 	j = nfs4_fl_calc_j_index(lseg, offset);
480 	idx = nfs4_fl_calc_ds_index(lseg, j);
481 	ds = nfs4_fl_prepare_ds(lseg, idx);
482 	if (!ds)
483 		return PNFS_NOT_ATTEMPTED;
484 
485 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
486 	if (IS_ERR(ds_clnt))
487 		return PNFS_NOT_ATTEMPTED;
488 
489 	dprintk("%s USE DS: %s cl_count %d\n", __func__,
490 		ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
491 
492 	/* No multipath support. Use first DS */
493 	atomic_inc(&ds->ds_clp->cl_count);
494 	hdr->ds_clp = ds->ds_clp;
495 	hdr->ds_commit_idx = idx;
496 	fh = nfs4_fl_select_ds_fh(lseg, j);
497 	if (fh)
498 		hdr->args.fh = fh;
499 
500 	hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
501 	hdr->mds_offset = offset;
502 
503 	/* Perform an asynchronous read to ds */
504 	nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
505 			  NFS_PROTO(hdr->inode), &filelayout_read_call_ops,
506 			  0, RPC_TASK_SOFTCONN);
507 	return PNFS_ATTEMPTED;
508 }
509 
510 /* Perform async writes. */
511 static enum pnfs_try_status
512 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
513 {
514 	struct pnfs_layout_segment *lseg = hdr->lseg;
515 	struct nfs4_pnfs_ds *ds;
516 	struct rpc_clnt *ds_clnt;
517 	loff_t offset = hdr->args.offset;
518 	u32 j, idx;
519 	struct nfs_fh *fh;
520 
521 	/* Retrieve the correct rpc_client for the byte range */
522 	j = nfs4_fl_calc_j_index(lseg, offset);
523 	idx = nfs4_fl_calc_ds_index(lseg, j);
524 	ds = nfs4_fl_prepare_ds(lseg, idx);
525 	if (!ds)
526 		return PNFS_NOT_ATTEMPTED;
527 
528 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
529 	if (IS_ERR(ds_clnt))
530 		return PNFS_NOT_ATTEMPTED;
531 
532 	dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
533 		__func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
534 		offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
535 
536 	hdr->pgio_done_cb = filelayout_write_done_cb;
537 	atomic_inc(&ds->ds_clp->cl_count);
538 	hdr->ds_clp = ds->ds_clp;
539 	hdr->ds_commit_idx = idx;
540 	fh = nfs4_fl_select_ds_fh(lseg, j);
541 	if (fh)
542 		hdr->args.fh = fh;
543 	hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
544 
545 	/* Perform an asynchronous write */
546 	nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
547 			  NFS_PROTO(hdr->inode), &filelayout_write_call_ops,
548 			  sync, RPC_TASK_SOFTCONN);
549 	return PNFS_ATTEMPTED;
550 }
551 
552 /*
553  * filelayout_check_layout()
554  *
555  * Make sure layout segment parameters are sane WRT the device.
556  * At this point no generic layer initialization of the lseg has occurred,
557  * and nothing has been added to the layout_hdr cache.
558  *
559  */
560 static int
561 filelayout_check_layout(struct pnfs_layout_hdr *lo,
562 			struct nfs4_filelayout_segment *fl,
563 			struct nfs4_layoutget_res *lgr,
564 			struct nfs4_deviceid *id,
565 			gfp_t gfp_flags)
566 {
567 	struct nfs4_deviceid_node *d;
568 	struct nfs4_file_layout_dsaddr *dsaddr;
569 	int status = -EINVAL;
570 
571 	dprintk("--> %s\n", __func__);
572 
573 	/* FIXME: remove this check when layout segment support is added */
574 	if (lgr->range.offset != 0 ||
575 	    lgr->range.length != NFS4_MAX_UINT64) {
576 		dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
577 			__func__);
578 		goto out;
579 	}
580 
581 	if (fl->pattern_offset > lgr->range.offset) {
582 		dprintk("%s pattern_offset %lld too large\n",
583 				__func__, fl->pattern_offset);
584 		goto out;
585 	}
586 
587 	if (!fl->stripe_unit) {
588 		dprintk("%s Invalid stripe unit (%u)\n",
589 			__func__, fl->stripe_unit);
590 		goto out;
591 	}
592 
593 	/* find and reference the deviceid */
594 	d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), id,
595 			lo->plh_lc_cred, gfp_flags);
596 	if (d == NULL)
597 		goto out;
598 
599 	dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
600 	/* Found deviceid is unavailable */
601 	if (filelayout_test_devid_unavailable(&dsaddr->id_node))
602 		goto out_put;
603 
604 	fl->dsaddr = dsaddr;
605 
606 	if (fl->first_stripe_index >= dsaddr->stripe_count) {
607 		dprintk("%s Bad first_stripe_index %u\n",
608 				__func__, fl->first_stripe_index);
609 		goto out_put;
610 	}
611 
612 	if ((fl->stripe_type == STRIPE_SPARSE &&
613 	    fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
614 	    (fl->stripe_type == STRIPE_DENSE &&
615 	    fl->num_fh != dsaddr->stripe_count)) {
616 		dprintk("%s num_fh %u not valid for given packing\n",
617 			__func__, fl->num_fh);
618 		goto out_put;
619 	}
620 
621 	status = 0;
622 out:
623 	dprintk("--> %s returns %d\n", __func__, status);
624 	return status;
625 out_put:
626 	nfs4_fl_put_deviceid(dsaddr);
627 	goto out;
628 }
629 
630 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
631 {
632 	int i;
633 
634 	for (i = 0; i < fl->num_fh; i++) {
635 		if (!fl->fh_array[i])
636 			break;
637 		kfree(fl->fh_array[i]);
638 	}
639 	kfree(fl->fh_array);
640 	fl->fh_array = NULL;
641 }
642 
643 static void
644 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
645 {
646 	filelayout_free_fh_array(fl);
647 	kfree(fl);
648 }
649 
650 static int
651 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
652 			 struct nfs4_filelayout_segment *fl,
653 			 struct nfs4_layoutget_res *lgr,
654 			 struct nfs4_deviceid *id,
655 			 gfp_t gfp_flags)
656 {
657 	struct xdr_stream stream;
658 	struct xdr_buf buf;
659 	struct page *scratch;
660 	__be32 *p;
661 	uint32_t nfl_util;
662 	int i;
663 
664 	dprintk("%s: set_layout_map Begin\n", __func__);
665 
666 	scratch = alloc_page(gfp_flags);
667 	if (!scratch)
668 		return -ENOMEM;
669 
670 	xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
671 	xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
672 
673 	/* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
674 	 * num_fh (4) */
675 	p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
676 	if (unlikely(!p))
677 		goto out_err;
678 
679 	memcpy(id, p, sizeof(*id));
680 	p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
681 	nfs4_print_deviceid(id);
682 
683 	nfl_util = be32_to_cpup(p++);
684 	if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
685 		fl->commit_through_mds = 1;
686 	if (nfl_util & NFL4_UFLG_DENSE)
687 		fl->stripe_type = STRIPE_DENSE;
688 	else
689 		fl->stripe_type = STRIPE_SPARSE;
690 	fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
691 
692 	fl->first_stripe_index = be32_to_cpup(p++);
693 	p = xdr_decode_hyper(p, &fl->pattern_offset);
694 	fl->num_fh = be32_to_cpup(p++);
695 
696 	dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
697 		__func__, nfl_util, fl->num_fh, fl->first_stripe_index,
698 		fl->pattern_offset);
699 
700 	/* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
701 	 * Futher checking is done in filelayout_check_layout */
702 	if (fl->num_fh >
703 	    max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
704 		goto out_err;
705 
706 	if (fl->num_fh > 0) {
707 		fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
708 				       gfp_flags);
709 		if (!fl->fh_array)
710 			goto out_err;
711 	}
712 
713 	for (i = 0; i < fl->num_fh; i++) {
714 		/* Do we want to use a mempool here? */
715 		fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
716 		if (!fl->fh_array[i])
717 			goto out_err_free;
718 
719 		p = xdr_inline_decode(&stream, 4);
720 		if (unlikely(!p))
721 			goto out_err_free;
722 		fl->fh_array[i]->size = be32_to_cpup(p++);
723 		if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
724 			printk(KERN_ERR "NFS: Too big fh %d received %d\n",
725 			       i, fl->fh_array[i]->size);
726 			goto out_err_free;
727 		}
728 
729 		p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
730 		if (unlikely(!p))
731 			goto out_err_free;
732 		memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
733 		dprintk("DEBUG: %s: fh len %d\n", __func__,
734 			fl->fh_array[i]->size);
735 	}
736 
737 	__free_page(scratch);
738 	return 0;
739 
740 out_err_free:
741 	filelayout_free_fh_array(fl);
742 out_err:
743 	__free_page(scratch);
744 	return -EIO;
745 }
746 
747 static void
748 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
749 {
750 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
751 
752 	dprintk("--> %s\n", __func__);
753 	nfs4_fl_put_deviceid(fl->dsaddr);
754 	/* This assumes a single RW lseg */
755 	if (lseg->pls_range.iomode == IOMODE_RW) {
756 		struct nfs4_filelayout *flo;
757 
758 		flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
759 		flo->commit_info.nbuckets = 0;
760 		kfree(flo->commit_info.buckets);
761 		flo->commit_info.buckets = NULL;
762 	}
763 	_filelayout_free_lseg(fl);
764 }
765 
766 static int
767 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
768 			     struct nfs_commit_info *cinfo,
769 			     gfp_t gfp_flags)
770 {
771 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
772 	struct pnfs_commit_bucket *buckets;
773 	int size, i;
774 
775 	if (fl->commit_through_mds)
776 		return 0;
777 
778 	size = (fl->stripe_type == STRIPE_SPARSE) ?
779 		fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
780 
781 	if (cinfo->ds->nbuckets >= size) {
782 		/* This assumes there is only one IOMODE_RW lseg.  What
783 		 * we really want to do is have a layout_hdr level
784 		 * dictionary of <multipath_list4, fh> keys, each
785 		 * associated with a struct list_head, populated by calls
786 		 * to filelayout_write_pagelist().
787 		 * */
788 		return 0;
789 	}
790 
791 	buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
792 			  gfp_flags);
793 	if (!buckets)
794 		return -ENOMEM;
795 	for (i = 0; i < size; i++) {
796 		INIT_LIST_HEAD(&buckets[i].written);
797 		INIT_LIST_HEAD(&buckets[i].committing);
798 		/* mark direct verifier as unset */
799 		buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
800 	}
801 
802 	spin_lock(cinfo->lock);
803 	if (cinfo->ds->nbuckets >= size)
804 		goto out;
805 	for (i = 0; i < cinfo->ds->nbuckets; i++) {
806 		list_splice(&cinfo->ds->buckets[i].written,
807 			    &buckets[i].written);
808 		list_splice(&cinfo->ds->buckets[i].committing,
809 			    &buckets[i].committing);
810 		buckets[i].direct_verf.committed =
811 			cinfo->ds->buckets[i].direct_verf.committed;
812 		buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
813 		buckets[i].clseg = cinfo->ds->buckets[i].clseg;
814 	}
815 	swap(cinfo->ds->buckets, buckets);
816 	cinfo->ds->nbuckets = size;
817 out:
818 	spin_unlock(cinfo->lock);
819 	kfree(buckets);
820 	return 0;
821 }
822 
823 static struct pnfs_layout_segment *
824 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
825 		      struct nfs4_layoutget_res *lgr,
826 		      gfp_t gfp_flags)
827 {
828 	struct nfs4_filelayout_segment *fl;
829 	int rc;
830 	struct nfs4_deviceid id;
831 
832 	dprintk("--> %s\n", __func__);
833 	fl = kzalloc(sizeof(*fl), gfp_flags);
834 	if (!fl)
835 		return NULL;
836 
837 	rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
838 	if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
839 		_filelayout_free_lseg(fl);
840 		return NULL;
841 	}
842 	return &fl->generic_hdr;
843 }
844 
845 /*
846  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
847  *
848  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
849  * of bytes (maximum @req->wb_bytes) that can be coalesced.
850  */
851 static size_t
852 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
853 		   struct nfs_page *req)
854 {
855 	unsigned int size;
856 	u64 p_stripe, r_stripe;
857 	u32 stripe_offset;
858 	u64 segment_offset = pgio->pg_lseg->pls_range.offset;
859 	u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
860 
861 	/* calls nfs_generic_pg_test */
862 	size = pnfs_generic_pg_test(pgio, prev, req);
863 	if (!size)
864 		return 0;
865 
866 	/* see if req and prev are in the same stripe */
867 	if (prev) {
868 		p_stripe = (u64)req_offset(prev) - segment_offset;
869 		r_stripe = (u64)req_offset(req) - segment_offset;
870 		do_div(p_stripe, stripe_unit);
871 		do_div(r_stripe, stripe_unit);
872 
873 		if (p_stripe != r_stripe)
874 			return 0;
875 	}
876 
877 	/* calculate remaining bytes in the current stripe */
878 	div_u64_rem((u64)req_offset(req) - segment_offset,
879 			stripe_unit,
880 			&stripe_offset);
881 	WARN_ON_ONCE(stripe_offset > stripe_unit);
882 	if (stripe_offset >= stripe_unit)
883 		return 0;
884 	return min(stripe_unit - (unsigned int)stripe_offset, size);
885 }
886 
887 static void
888 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
889 			struct nfs_page *req)
890 {
891 	if (!pgio->pg_lseg)
892 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
893 					   req->wb_context,
894 					   0,
895 					   NFS4_MAX_UINT64,
896 					   IOMODE_READ,
897 					   GFP_KERNEL);
898 	/* If no lseg, fall back to read through mds */
899 	if (pgio->pg_lseg == NULL)
900 		nfs_pageio_reset_read_mds(pgio);
901 }
902 
903 static void
904 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
905 			 struct nfs_page *req)
906 {
907 	struct nfs_commit_info cinfo;
908 	int status;
909 
910 	if (!pgio->pg_lseg)
911 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
912 					   req->wb_context,
913 					   0,
914 					   NFS4_MAX_UINT64,
915 					   IOMODE_RW,
916 					   GFP_NOFS);
917 	/* If no lseg, fall back to write through mds */
918 	if (pgio->pg_lseg == NULL)
919 		goto out_mds;
920 	nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
921 	status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
922 	if (status < 0) {
923 		pnfs_put_lseg(pgio->pg_lseg);
924 		pgio->pg_lseg = NULL;
925 		goto out_mds;
926 	}
927 	return;
928 out_mds:
929 	nfs_pageio_reset_write_mds(pgio);
930 }
931 
932 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
933 	.pg_init = filelayout_pg_init_read,
934 	.pg_test = filelayout_pg_test,
935 	.pg_doio = pnfs_generic_pg_readpages,
936 	.pg_cleanup = pnfs_generic_pg_cleanup,
937 };
938 
939 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
940 	.pg_init = filelayout_pg_init_write,
941 	.pg_test = filelayout_pg_test,
942 	.pg_doio = pnfs_generic_pg_writepages,
943 	.pg_cleanup = pnfs_generic_pg_cleanup,
944 };
945 
946 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
947 {
948 	if (fl->stripe_type == STRIPE_SPARSE)
949 		return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
950 	else
951 		return j;
952 }
953 
954 static void
955 filelayout_mark_request_commit(struct nfs_page *req,
956 			       struct pnfs_layout_segment *lseg,
957 			       struct nfs_commit_info *cinfo,
958 			       u32 ds_commit_idx)
959 
960 {
961 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
962 	u32 i, j;
963 
964 	if (fl->commit_through_mds) {
965 		nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
966 	} else {
967 		/* Note that we are calling nfs4_fl_calc_j_index on each page
968 		 * that ends up being committed to a data server.  An attractive
969 		 * alternative is to add a field to nfs_write_data and nfs_page
970 		 * to store the value calculated in filelayout_write_pagelist
971 		 * and just use that here.
972 		 */
973 		j = nfs4_fl_calc_j_index(lseg, req_offset(req));
974 		i = select_bucket_index(fl, j);
975 		pnfs_layout_mark_request_commit(req, lseg, cinfo, i);
976 	}
977 }
978 
979 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
980 {
981 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
982 
983 	if (flseg->stripe_type == STRIPE_SPARSE)
984 		return i;
985 	else
986 		return nfs4_fl_calc_ds_index(lseg, i);
987 }
988 
989 static struct nfs_fh *
990 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
991 {
992 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
993 
994 	if (flseg->stripe_type == STRIPE_SPARSE) {
995 		if (flseg->num_fh == 1)
996 			i = 0;
997 		else if (flseg->num_fh == 0)
998 			/* Use the MDS OPEN fh set in nfs_read_rpcsetup */
999 			return NULL;
1000 	}
1001 	return flseg->fh_array[i];
1002 }
1003 
1004 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1005 {
1006 	struct pnfs_layout_segment *lseg = data->lseg;
1007 	struct nfs4_pnfs_ds *ds;
1008 	struct rpc_clnt *ds_clnt;
1009 	u32 idx;
1010 	struct nfs_fh *fh;
1011 
1012 	idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1013 	ds = nfs4_fl_prepare_ds(lseg, idx);
1014 	if (!ds)
1015 		goto out_err;
1016 
1017 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1018 	if (IS_ERR(ds_clnt))
1019 		goto out_err;
1020 
1021 	dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1022 		data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1023 	data->commit_done_cb = filelayout_commit_done_cb;
1024 	atomic_inc(&ds->ds_clp->cl_count);
1025 	data->ds_clp = ds->ds_clp;
1026 	fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1027 	if (fh)
1028 		data->args.fh = fh;
1029 	return nfs_initiate_commit(ds_clnt, data, NFS_PROTO(data->inode),
1030 				   &filelayout_commit_call_ops, how,
1031 				   RPC_TASK_SOFTCONN);
1032 out_err:
1033 	pnfs_generic_prepare_to_resend_writes(data);
1034 	pnfs_generic_commit_release(data);
1035 	return -EAGAIN;
1036 }
1037 
1038 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1039  *				   for @page
1040  * @cinfo - commit info for current inode
1041  * @page - page to search for matching head request
1042  *
1043  * Returns a the head request if one is found, otherwise returns NULL.
1044  */
1045 static struct nfs_page *
1046 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1047 {
1048 	struct nfs_page *freq, *t;
1049 	struct pnfs_commit_bucket *b;
1050 	int i;
1051 
1052 	/* Linearly search the commit lists for each bucket until a matching
1053 	 * request is found */
1054 	for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1055 		list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1056 			if (freq->wb_page == page)
1057 				return freq->wb_head;
1058 		}
1059 		list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1060 			if (freq->wb_page == page)
1061 				return freq->wb_head;
1062 		}
1063 	}
1064 
1065 	return NULL;
1066 }
1067 
1068 static int
1069 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1070 			   int how, struct nfs_commit_info *cinfo)
1071 {
1072 	return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1073 					    filelayout_initiate_commit);
1074 }
1075 
1076 static struct nfs4_deviceid_node *
1077 filelayout_alloc_deviceid_node(struct nfs_server *server,
1078 		struct pnfs_device *pdev, gfp_t gfp_flags)
1079 {
1080 	struct nfs4_file_layout_dsaddr *dsaddr;
1081 
1082 	dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1083 	if (!dsaddr)
1084 		return NULL;
1085 	return &dsaddr->id_node;
1086 }
1087 
1088 static void
1089 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1090 {
1091 	nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1092 }
1093 
1094 static struct pnfs_layout_hdr *
1095 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1096 {
1097 	struct nfs4_filelayout *flo;
1098 
1099 	flo = kzalloc(sizeof(*flo), gfp_flags);
1100 	return flo != NULL ? &flo->generic_hdr : NULL;
1101 }
1102 
1103 static void
1104 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1105 {
1106 	kfree(FILELAYOUT_FROM_HDR(lo));
1107 }
1108 
1109 static struct pnfs_ds_commit_info *
1110 filelayout_get_ds_info(struct inode *inode)
1111 {
1112 	struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1113 
1114 	if (layout == NULL)
1115 		return NULL;
1116 	else
1117 		return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1118 }
1119 
1120 static struct pnfs_layoutdriver_type filelayout_type = {
1121 	.id			= LAYOUT_NFSV4_1_FILES,
1122 	.name			= "LAYOUT_NFSV4_1_FILES",
1123 	.owner			= THIS_MODULE,
1124 	.alloc_layout_hdr	= filelayout_alloc_layout_hdr,
1125 	.free_layout_hdr	= filelayout_free_layout_hdr,
1126 	.alloc_lseg		= filelayout_alloc_lseg,
1127 	.free_lseg		= filelayout_free_lseg,
1128 	.pg_read_ops		= &filelayout_pg_read_ops,
1129 	.pg_write_ops		= &filelayout_pg_write_ops,
1130 	.get_ds_info		= &filelayout_get_ds_info,
1131 	.mark_request_commit	= filelayout_mark_request_commit,
1132 	.clear_request_commit	= pnfs_generic_clear_request_commit,
1133 	.scan_commit_lists	= pnfs_generic_scan_commit_lists,
1134 	.recover_commit_reqs	= pnfs_generic_recover_commit_reqs,
1135 	.search_commit_reqs	= filelayout_search_commit_reqs,
1136 	.commit_pagelist	= filelayout_commit_pagelist,
1137 	.read_pagelist		= filelayout_read_pagelist,
1138 	.write_pagelist		= filelayout_write_pagelist,
1139 	.alloc_deviceid_node	= filelayout_alloc_deviceid_node,
1140 	.free_deviceid_node	= filelayout_free_deveiceid_node,
1141 };
1142 
1143 static int __init nfs4filelayout_init(void)
1144 {
1145 	printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1146 	       __func__);
1147 	return pnfs_register_layoutdriver(&filelayout_type);
1148 }
1149 
1150 static void __exit nfs4filelayout_exit(void)
1151 {
1152 	printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1153 	       __func__);
1154 	pnfs_unregister_layoutdriver(&filelayout_type);
1155 }
1156 
1157 MODULE_ALIAS("nfs-layouttype4-1");
1158 
1159 module_init(nfs4filelayout_init);
1160 module_exit(nfs4filelayout_exit);
1161