xref: /openbmc/linux/fs/nfs/filelayout/filelayout.c (revision e2f1cf25)
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 		rpc_wake_up(&tbl->slot_tbl_waitq);
206 		/* fall through */
207 	default:
208 reset:
209 		dprintk("%s Retry through MDS. Error %d\n", __func__,
210 			task->tk_status);
211 		return -NFS4ERR_RESET_TO_MDS;
212 	}
213 out:
214 	task->tk_status = 0;
215 	return -EAGAIN;
216 out_bad_stateid:
217 	task->tk_status = -EIO;
218 	return 0;
219 wait_on_recovery:
220 	rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
221 	if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
222 		rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
223 	goto out;
224 }
225 
226 /* NFS_PROTO call done callback routines */
227 
228 static int filelayout_read_done_cb(struct rpc_task *task,
229 				struct nfs_pgio_header *hdr)
230 {
231 	int err;
232 
233 	trace_nfs4_pnfs_read(hdr, task->tk_status);
234 	err = filelayout_async_handle_error(task, hdr->args.context->state,
235 					    hdr->ds_clp, hdr->lseg);
236 
237 	switch (err) {
238 	case -NFS4ERR_RESET_TO_MDS:
239 		filelayout_reset_read(hdr);
240 		return task->tk_status;
241 	case -EAGAIN:
242 		rpc_restart_call_prepare(task);
243 		return -EAGAIN;
244 	}
245 
246 	return 0;
247 }
248 
249 /*
250  * We reference the rpc_cred of the first WRITE that triggers the need for
251  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
252  * rfc5661 is not clear about which credential should be used.
253  */
254 static void
255 filelayout_set_layoutcommit(struct nfs_pgio_header *hdr)
256 {
257 
258 	if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
259 	    hdr->res.verf->committed != NFS_DATA_SYNC)
260 		return;
261 
262 	pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
263 			hdr->mds_offset + hdr->res.count);
264 	dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
265 		(unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
266 }
267 
268 bool
269 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
270 {
271 	return filelayout_test_devid_invalid(node) ||
272 		nfs4_test_deviceid_unavailable(node);
273 }
274 
275 static bool
276 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
277 {
278 	struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
279 
280 	return filelayout_test_devid_unavailable(node);
281 }
282 
283 /*
284  * Call ops for the async read/write cases
285  * In the case of dense layouts, the offset needs to be reset to its
286  * original value.
287  */
288 static void filelayout_read_prepare(struct rpc_task *task, void *data)
289 {
290 	struct nfs_pgio_header *hdr = data;
291 
292 	if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
293 		rpc_exit(task, -EIO);
294 		return;
295 	}
296 	if (filelayout_reset_to_mds(hdr->lseg)) {
297 		dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
298 		filelayout_reset_read(hdr);
299 		rpc_exit(task, 0);
300 		return;
301 	}
302 	hdr->pgio_done_cb = filelayout_read_done_cb;
303 
304 	if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
305 			&hdr->args.seq_args,
306 			&hdr->res.seq_res,
307 			task))
308 		return;
309 	if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
310 			hdr->args.lock_context, FMODE_READ) == -EIO)
311 		rpc_exit(task, -EIO); /* lost lock, terminate I/O */
312 }
313 
314 static void filelayout_read_call_done(struct rpc_task *task, void *data)
315 {
316 	struct nfs_pgio_header *hdr = data;
317 
318 	dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
319 
320 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
321 	    task->tk_status == 0) {
322 		nfs41_sequence_done(task, &hdr->res.seq_res);
323 		return;
324 	}
325 
326 	/* Note this may cause RPC to be resent */
327 	hdr->mds_ops->rpc_call_done(task, data);
328 }
329 
330 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
331 {
332 	struct nfs_pgio_header *hdr = data;
333 
334 	rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
335 }
336 
337 static int filelayout_write_done_cb(struct rpc_task *task,
338 				struct nfs_pgio_header *hdr)
339 {
340 	int err;
341 
342 	trace_nfs4_pnfs_write(hdr, task->tk_status);
343 	err = filelayout_async_handle_error(task, hdr->args.context->state,
344 					    hdr->ds_clp, hdr->lseg);
345 
346 	switch (err) {
347 	case -NFS4ERR_RESET_TO_MDS:
348 		filelayout_reset_write(hdr);
349 		return task->tk_status;
350 	case -EAGAIN:
351 		rpc_restart_call_prepare(task);
352 		return -EAGAIN;
353 	}
354 
355 	filelayout_set_layoutcommit(hdr);
356 	return 0;
357 }
358 
359 static int filelayout_commit_done_cb(struct rpc_task *task,
360 				     struct nfs_commit_data *data)
361 {
362 	int err;
363 
364 	trace_nfs4_pnfs_commit_ds(data, task->tk_status);
365 	err = filelayout_async_handle_error(task, NULL, data->ds_clp,
366 					    data->lseg);
367 
368 	switch (err) {
369 	case -NFS4ERR_RESET_TO_MDS:
370 		pnfs_generic_prepare_to_resend_writes(data);
371 		return -EAGAIN;
372 	case -EAGAIN:
373 		rpc_restart_call_prepare(task);
374 		return -EAGAIN;
375 	}
376 
377 	if (data->verf.committed == NFS_UNSTABLE)
378 		pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
379 
380 	return 0;
381 }
382 
383 static void filelayout_write_prepare(struct rpc_task *task, void *data)
384 {
385 	struct nfs_pgio_header *hdr = data;
386 
387 	if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
388 		rpc_exit(task, -EIO);
389 		return;
390 	}
391 	if (filelayout_reset_to_mds(hdr->lseg)) {
392 		dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
393 		filelayout_reset_write(hdr);
394 		rpc_exit(task, 0);
395 		return;
396 	}
397 	if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
398 			&hdr->args.seq_args,
399 			&hdr->res.seq_res,
400 			task))
401 		return;
402 	if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
403 			hdr->args.lock_context, FMODE_WRITE) == -EIO)
404 		rpc_exit(task, -EIO); /* lost lock, terminate I/O */
405 }
406 
407 static void filelayout_write_call_done(struct rpc_task *task, void *data)
408 {
409 	struct nfs_pgio_header *hdr = data;
410 
411 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
412 	    task->tk_status == 0) {
413 		nfs41_sequence_done(task, &hdr->res.seq_res);
414 		return;
415 	}
416 
417 	/* Note this may cause RPC to be resent */
418 	hdr->mds_ops->rpc_call_done(task, data);
419 }
420 
421 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
422 {
423 	struct nfs_pgio_header *hdr = data;
424 
425 	rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
426 }
427 
428 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
429 {
430 	struct nfs_commit_data *wdata = data;
431 
432 	nfs41_setup_sequence(wdata->ds_clp->cl_session,
433 			&wdata->args.seq_args,
434 			&wdata->res.seq_res,
435 			task);
436 }
437 
438 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
439 {
440 	struct nfs_commit_data *cdata = data;
441 
442 	rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
443 }
444 
445 static const struct rpc_call_ops filelayout_read_call_ops = {
446 	.rpc_call_prepare = filelayout_read_prepare,
447 	.rpc_call_done = filelayout_read_call_done,
448 	.rpc_count_stats = filelayout_read_count_stats,
449 	.rpc_release = pnfs_generic_rw_release,
450 };
451 
452 static const struct rpc_call_ops filelayout_write_call_ops = {
453 	.rpc_call_prepare = filelayout_write_prepare,
454 	.rpc_call_done = filelayout_write_call_done,
455 	.rpc_count_stats = filelayout_write_count_stats,
456 	.rpc_release = pnfs_generic_rw_release,
457 };
458 
459 static const struct rpc_call_ops filelayout_commit_call_ops = {
460 	.rpc_call_prepare = filelayout_commit_prepare,
461 	.rpc_call_done = pnfs_generic_write_commit_done,
462 	.rpc_count_stats = filelayout_commit_count_stats,
463 	.rpc_release = pnfs_generic_commit_release,
464 };
465 
466 static enum pnfs_try_status
467 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
468 {
469 	struct pnfs_layout_segment *lseg = hdr->lseg;
470 	struct nfs4_pnfs_ds *ds;
471 	struct rpc_clnt *ds_clnt;
472 	loff_t offset = hdr->args.offset;
473 	u32 j, idx;
474 	struct nfs_fh *fh;
475 
476 	dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
477 		__func__, hdr->inode->i_ino,
478 		hdr->args.pgbase, (size_t)hdr->args.count, offset);
479 
480 	/* Retrieve the correct rpc_client for the byte range */
481 	j = nfs4_fl_calc_j_index(lseg, offset);
482 	idx = nfs4_fl_calc_ds_index(lseg, j);
483 	ds = nfs4_fl_prepare_ds(lseg, idx);
484 	if (!ds)
485 		return PNFS_NOT_ATTEMPTED;
486 
487 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
488 	if (IS_ERR(ds_clnt))
489 		return PNFS_NOT_ATTEMPTED;
490 
491 	dprintk("%s USE DS: %s cl_count %d\n", __func__,
492 		ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
493 
494 	/* No multipath support. Use first DS */
495 	atomic_inc(&ds->ds_clp->cl_count);
496 	hdr->ds_clp = ds->ds_clp;
497 	hdr->ds_commit_idx = idx;
498 	fh = nfs4_fl_select_ds_fh(lseg, j);
499 	if (fh)
500 		hdr->args.fh = fh;
501 
502 	hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
503 	hdr->mds_offset = offset;
504 
505 	/* Perform an asynchronous read to ds */
506 	nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
507 			  NFS_PROTO(hdr->inode), &filelayout_read_call_ops,
508 			  0, RPC_TASK_SOFTCONN);
509 	return PNFS_ATTEMPTED;
510 }
511 
512 /* Perform async writes. */
513 static enum pnfs_try_status
514 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
515 {
516 	struct pnfs_layout_segment *lseg = hdr->lseg;
517 	struct nfs4_pnfs_ds *ds;
518 	struct rpc_clnt *ds_clnt;
519 	loff_t offset = hdr->args.offset;
520 	u32 j, idx;
521 	struct nfs_fh *fh;
522 
523 	/* Retrieve the correct rpc_client for the byte range */
524 	j = nfs4_fl_calc_j_index(lseg, offset);
525 	idx = nfs4_fl_calc_ds_index(lseg, j);
526 	ds = nfs4_fl_prepare_ds(lseg, idx);
527 	if (!ds)
528 		return PNFS_NOT_ATTEMPTED;
529 
530 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
531 	if (IS_ERR(ds_clnt))
532 		return PNFS_NOT_ATTEMPTED;
533 
534 	dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
535 		__func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
536 		offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
537 
538 	hdr->pgio_done_cb = filelayout_write_done_cb;
539 	atomic_inc(&ds->ds_clp->cl_count);
540 	hdr->ds_clp = ds->ds_clp;
541 	hdr->ds_commit_idx = idx;
542 	fh = nfs4_fl_select_ds_fh(lseg, j);
543 	if (fh)
544 		hdr->args.fh = fh;
545 	hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
546 
547 	/* Perform an asynchronous write */
548 	nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
549 			  NFS_PROTO(hdr->inode), &filelayout_write_call_ops,
550 			  sync, RPC_TASK_SOFTCONN);
551 	return PNFS_ATTEMPTED;
552 }
553 
554 /*
555  * filelayout_check_layout()
556  *
557  * Make sure layout segment parameters are sane WRT the device.
558  * At this point no generic layer initialization of the lseg has occurred,
559  * and nothing has been added to the layout_hdr cache.
560  *
561  */
562 static int
563 filelayout_check_layout(struct pnfs_layout_hdr *lo,
564 			struct nfs4_filelayout_segment *fl,
565 			struct nfs4_layoutget_res *lgr,
566 			struct nfs4_deviceid *id,
567 			gfp_t gfp_flags)
568 {
569 	struct nfs4_deviceid_node *d;
570 	struct nfs4_file_layout_dsaddr *dsaddr;
571 	int status = -EINVAL;
572 
573 	dprintk("--> %s\n", __func__);
574 
575 	/* FIXME: remove this check when layout segment support is added */
576 	if (lgr->range.offset != 0 ||
577 	    lgr->range.length != NFS4_MAX_UINT64) {
578 		dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
579 			__func__);
580 		goto out;
581 	}
582 
583 	if (fl->pattern_offset > lgr->range.offset) {
584 		dprintk("%s pattern_offset %lld too large\n",
585 				__func__, fl->pattern_offset);
586 		goto out;
587 	}
588 
589 	if (!fl->stripe_unit) {
590 		dprintk("%s Invalid stripe unit (%u)\n",
591 			__func__, fl->stripe_unit);
592 		goto out;
593 	}
594 
595 	/* find and reference the deviceid */
596 	d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), id,
597 			lo->plh_lc_cred, gfp_flags);
598 	if (d == NULL)
599 		goto out;
600 
601 	dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
602 	/* Found deviceid is unavailable */
603 	if (filelayout_test_devid_unavailable(&dsaddr->id_node))
604 		goto out_put;
605 
606 	fl->dsaddr = dsaddr;
607 
608 	if (fl->first_stripe_index >= dsaddr->stripe_count) {
609 		dprintk("%s Bad first_stripe_index %u\n",
610 				__func__, fl->first_stripe_index);
611 		goto out_put;
612 	}
613 
614 	if ((fl->stripe_type == STRIPE_SPARSE &&
615 	    fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
616 	    (fl->stripe_type == STRIPE_DENSE &&
617 	    fl->num_fh != dsaddr->stripe_count)) {
618 		dprintk("%s num_fh %u not valid for given packing\n",
619 			__func__, fl->num_fh);
620 		goto out_put;
621 	}
622 
623 	status = 0;
624 out:
625 	dprintk("--> %s returns %d\n", __func__, status);
626 	return status;
627 out_put:
628 	nfs4_fl_put_deviceid(dsaddr);
629 	goto out;
630 }
631 
632 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
633 {
634 	int i;
635 
636 	for (i = 0; i < fl->num_fh; i++) {
637 		if (!fl->fh_array[i])
638 			break;
639 		kfree(fl->fh_array[i]);
640 	}
641 	kfree(fl->fh_array);
642 	fl->fh_array = NULL;
643 }
644 
645 static void
646 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
647 {
648 	filelayout_free_fh_array(fl);
649 	kfree(fl);
650 }
651 
652 static int
653 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
654 			 struct nfs4_filelayout_segment *fl,
655 			 struct nfs4_layoutget_res *lgr,
656 			 struct nfs4_deviceid *id,
657 			 gfp_t gfp_flags)
658 {
659 	struct xdr_stream stream;
660 	struct xdr_buf buf;
661 	struct page *scratch;
662 	__be32 *p;
663 	uint32_t nfl_util;
664 	int i;
665 
666 	dprintk("%s: set_layout_map Begin\n", __func__);
667 
668 	scratch = alloc_page(gfp_flags);
669 	if (!scratch)
670 		return -ENOMEM;
671 
672 	xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
673 	xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
674 
675 	/* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
676 	 * num_fh (4) */
677 	p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
678 	if (unlikely(!p))
679 		goto out_err;
680 
681 	memcpy(id, p, sizeof(*id));
682 	p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
683 	nfs4_print_deviceid(id);
684 
685 	nfl_util = be32_to_cpup(p++);
686 	if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
687 		fl->commit_through_mds = 1;
688 	if (nfl_util & NFL4_UFLG_DENSE)
689 		fl->stripe_type = STRIPE_DENSE;
690 	else
691 		fl->stripe_type = STRIPE_SPARSE;
692 	fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
693 
694 	fl->first_stripe_index = be32_to_cpup(p++);
695 	p = xdr_decode_hyper(p, &fl->pattern_offset);
696 	fl->num_fh = be32_to_cpup(p++);
697 
698 	dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
699 		__func__, nfl_util, fl->num_fh, fl->first_stripe_index,
700 		fl->pattern_offset);
701 
702 	/* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
703 	 * Futher checking is done in filelayout_check_layout */
704 	if (fl->num_fh >
705 	    max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
706 		goto out_err;
707 
708 	if (fl->num_fh > 0) {
709 		fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
710 				       gfp_flags);
711 		if (!fl->fh_array)
712 			goto out_err;
713 	}
714 
715 	for (i = 0; i < fl->num_fh; i++) {
716 		/* Do we want to use a mempool here? */
717 		fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
718 		if (!fl->fh_array[i])
719 			goto out_err_free;
720 
721 		p = xdr_inline_decode(&stream, 4);
722 		if (unlikely(!p))
723 			goto out_err_free;
724 		fl->fh_array[i]->size = be32_to_cpup(p++);
725 		if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
726 			printk(KERN_ERR "NFS: Too big fh %d received %d\n",
727 			       i, fl->fh_array[i]->size);
728 			goto out_err_free;
729 		}
730 
731 		p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
732 		if (unlikely(!p))
733 			goto out_err_free;
734 		memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
735 		dprintk("DEBUG: %s: fh len %d\n", __func__,
736 			fl->fh_array[i]->size);
737 	}
738 
739 	__free_page(scratch);
740 	return 0;
741 
742 out_err_free:
743 	filelayout_free_fh_array(fl);
744 out_err:
745 	__free_page(scratch);
746 	return -EIO;
747 }
748 
749 static void
750 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
751 {
752 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
753 
754 	dprintk("--> %s\n", __func__);
755 	nfs4_fl_put_deviceid(fl->dsaddr);
756 	/* This assumes a single RW lseg */
757 	if (lseg->pls_range.iomode == IOMODE_RW) {
758 		struct nfs4_filelayout *flo;
759 
760 		flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
761 		flo->commit_info.nbuckets = 0;
762 		kfree(flo->commit_info.buckets);
763 		flo->commit_info.buckets = NULL;
764 	}
765 	_filelayout_free_lseg(fl);
766 }
767 
768 static int
769 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
770 			     struct nfs_commit_info *cinfo,
771 			     gfp_t gfp_flags)
772 {
773 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
774 	struct pnfs_commit_bucket *buckets;
775 	int size, i;
776 
777 	if (fl->commit_through_mds)
778 		return 0;
779 
780 	size = (fl->stripe_type == STRIPE_SPARSE) ?
781 		fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
782 
783 	if (cinfo->ds->nbuckets >= size) {
784 		/* This assumes there is only one IOMODE_RW lseg.  What
785 		 * we really want to do is have a layout_hdr level
786 		 * dictionary of <multipath_list4, fh> keys, each
787 		 * associated with a struct list_head, populated by calls
788 		 * to filelayout_write_pagelist().
789 		 * */
790 		return 0;
791 	}
792 
793 	buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
794 			  gfp_flags);
795 	if (!buckets)
796 		return -ENOMEM;
797 	for (i = 0; i < size; i++) {
798 		INIT_LIST_HEAD(&buckets[i].written);
799 		INIT_LIST_HEAD(&buckets[i].committing);
800 		/* mark direct verifier as unset */
801 		buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
802 	}
803 
804 	spin_lock(cinfo->lock);
805 	if (cinfo->ds->nbuckets >= size)
806 		goto out;
807 	for (i = 0; i < cinfo->ds->nbuckets; i++) {
808 		list_splice(&cinfo->ds->buckets[i].written,
809 			    &buckets[i].written);
810 		list_splice(&cinfo->ds->buckets[i].committing,
811 			    &buckets[i].committing);
812 		buckets[i].direct_verf.committed =
813 			cinfo->ds->buckets[i].direct_verf.committed;
814 		buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
815 		buckets[i].clseg = cinfo->ds->buckets[i].clseg;
816 	}
817 	swap(cinfo->ds->buckets, buckets);
818 	cinfo->ds->nbuckets = size;
819 out:
820 	spin_unlock(cinfo->lock);
821 	kfree(buckets);
822 	return 0;
823 }
824 
825 static struct pnfs_layout_segment *
826 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
827 		      struct nfs4_layoutget_res *lgr,
828 		      gfp_t gfp_flags)
829 {
830 	struct nfs4_filelayout_segment *fl;
831 	int rc;
832 	struct nfs4_deviceid id;
833 
834 	dprintk("--> %s\n", __func__);
835 	fl = kzalloc(sizeof(*fl), gfp_flags);
836 	if (!fl)
837 		return NULL;
838 
839 	rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
840 	if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
841 		_filelayout_free_lseg(fl);
842 		return NULL;
843 	}
844 	return &fl->generic_hdr;
845 }
846 
847 /*
848  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
849  *
850  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
851  * of bytes (maximum @req->wb_bytes) that can be coalesced.
852  */
853 static size_t
854 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
855 		   struct nfs_page *req)
856 {
857 	unsigned int size;
858 	u64 p_stripe, r_stripe;
859 	u32 stripe_offset;
860 	u64 segment_offset = pgio->pg_lseg->pls_range.offset;
861 	u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
862 
863 	/* calls nfs_generic_pg_test */
864 	size = pnfs_generic_pg_test(pgio, prev, req);
865 	if (!size)
866 		return 0;
867 
868 	/* see if req and prev are in the same stripe */
869 	if (prev) {
870 		p_stripe = (u64)req_offset(prev) - segment_offset;
871 		r_stripe = (u64)req_offset(req) - segment_offset;
872 		do_div(p_stripe, stripe_unit);
873 		do_div(r_stripe, stripe_unit);
874 
875 		if (p_stripe != r_stripe)
876 			return 0;
877 	}
878 
879 	/* calculate remaining bytes in the current stripe */
880 	div_u64_rem((u64)req_offset(req) - segment_offset,
881 			stripe_unit,
882 			&stripe_offset);
883 	WARN_ON_ONCE(stripe_offset > stripe_unit);
884 	if (stripe_offset >= stripe_unit)
885 		return 0;
886 	return min(stripe_unit - (unsigned int)stripe_offset, size);
887 }
888 
889 static void
890 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
891 			struct nfs_page *req)
892 {
893 	if (!pgio->pg_lseg)
894 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
895 					   req->wb_context,
896 					   0,
897 					   NFS4_MAX_UINT64,
898 					   IOMODE_READ,
899 					   GFP_KERNEL);
900 	/* If no lseg, fall back to read through mds */
901 	if (pgio->pg_lseg == NULL)
902 		nfs_pageio_reset_read_mds(pgio);
903 }
904 
905 static void
906 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
907 			 struct nfs_page *req)
908 {
909 	struct nfs_commit_info cinfo;
910 	int status;
911 
912 	if (!pgio->pg_lseg)
913 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
914 					   req->wb_context,
915 					   0,
916 					   NFS4_MAX_UINT64,
917 					   IOMODE_RW,
918 					   GFP_NOFS);
919 	/* If no lseg, fall back to write through mds */
920 	if (pgio->pg_lseg == NULL)
921 		goto out_mds;
922 	nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
923 	status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
924 	if (status < 0) {
925 		pnfs_put_lseg(pgio->pg_lseg);
926 		pgio->pg_lseg = NULL;
927 		goto out_mds;
928 	}
929 	return;
930 out_mds:
931 	nfs_pageio_reset_write_mds(pgio);
932 }
933 
934 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
935 	.pg_init = filelayout_pg_init_read,
936 	.pg_test = filelayout_pg_test,
937 	.pg_doio = pnfs_generic_pg_readpages,
938 	.pg_cleanup = pnfs_generic_pg_cleanup,
939 };
940 
941 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
942 	.pg_init = filelayout_pg_init_write,
943 	.pg_test = filelayout_pg_test,
944 	.pg_doio = pnfs_generic_pg_writepages,
945 	.pg_cleanup = pnfs_generic_pg_cleanup,
946 };
947 
948 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
949 {
950 	if (fl->stripe_type == STRIPE_SPARSE)
951 		return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
952 	else
953 		return j;
954 }
955 
956 static void
957 filelayout_mark_request_commit(struct nfs_page *req,
958 			       struct pnfs_layout_segment *lseg,
959 			       struct nfs_commit_info *cinfo,
960 			       u32 ds_commit_idx)
961 
962 {
963 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
964 	u32 i, j;
965 
966 	if (fl->commit_through_mds) {
967 		nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
968 	} else {
969 		/* Note that we are calling nfs4_fl_calc_j_index on each page
970 		 * that ends up being committed to a data server.  An attractive
971 		 * alternative is to add a field to nfs_write_data and nfs_page
972 		 * to store the value calculated in filelayout_write_pagelist
973 		 * and just use that here.
974 		 */
975 		j = nfs4_fl_calc_j_index(lseg, req_offset(req));
976 		i = select_bucket_index(fl, j);
977 		pnfs_layout_mark_request_commit(req, lseg, cinfo, i);
978 	}
979 }
980 
981 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
982 {
983 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
984 
985 	if (flseg->stripe_type == STRIPE_SPARSE)
986 		return i;
987 	else
988 		return nfs4_fl_calc_ds_index(lseg, i);
989 }
990 
991 static struct nfs_fh *
992 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
993 {
994 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
995 
996 	if (flseg->stripe_type == STRIPE_SPARSE) {
997 		if (flseg->num_fh == 1)
998 			i = 0;
999 		else if (flseg->num_fh == 0)
1000 			/* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1001 			return NULL;
1002 	}
1003 	return flseg->fh_array[i];
1004 }
1005 
1006 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1007 {
1008 	struct pnfs_layout_segment *lseg = data->lseg;
1009 	struct nfs4_pnfs_ds *ds;
1010 	struct rpc_clnt *ds_clnt;
1011 	u32 idx;
1012 	struct nfs_fh *fh;
1013 
1014 	idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1015 	ds = nfs4_fl_prepare_ds(lseg, idx);
1016 	if (!ds)
1017 		goto out_err;
1018 
1019 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1020 	if (IS_ERR(ds_clnt))
1021 		goto out_err;
1022 
1023 	dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1024 		data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1025 	data->commit_done_cb = filelayout_commit_done_cb;
1026 	atomic_inc(&ds->ds_clp->cl_count);
1027 	data->ds_clp = ds->ds_clp;
1028 	fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1029 	if (fh)
1030 		data->args.fh = fh;
1031 	return nfs_initiate_commit(ds_clnt, data, NFS_PROTO(data->inode),
1032 				   &filelayout_commit_call_ops, how,
1033 				   RPC_TASK_SOFTCONN);
1034 out_err:
1035 	pnfs_generic_prepare_to_resend_writes(data);
1036 	pnfs_generic_commit_release(data);
1037 	return -EAGAIN;
1038 }
1039 
1040 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1041  *				   for @page
1042  * @cinfo - commit info for current inode
1043  * @page - page to search for matching head request
1044  *
1045  * Returns a the head request if one is found, otherwise returns NULL.
1046  */
1047 static struct nfs_page *
1048 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1049 {
1050 	struct nfs_page *freq, *t;
1051 	struct pnfs_commit_bucket *b;
1052 	int i;
1053 
1054 	/* Linearly search the commit lists for each bucket until a matching
1055 	 * request is found */
1056 	for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1057 		list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1058 			if (freq->wb_page == page)
1059 				return freq->wb_head;
1060 		}
1061 		list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1062 			if (freq->wb_page == page)
1063 				return freq->wb_head;
1064 		}
1065 	}
1066 
1067 	return NULL;
1068 }
1069 
1070 static int
1071 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1072 			   int how, struct nfs_commit_info *cinfo)
1073 {
1074 	return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1075 					    filelayout_initiate_commit);
1076 }
1077 
1078 static struct nfs4_deviceid_node *
1079 filelayout_alloc_deviceid_node(struct nfs_server *server,
1080 		struct pnfs_device *pdev, gfp_t gfp_flags)
1081 {
1082 	struct nfs4_file_layout_dsaddr *dsaddr;
1083 
1084 	dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1085 	if (!dsaddr)
1086 		return NULL;
1087 	return &dsaddr->id_node;
1088 }
1089 
1090 static void
1091 filelayout_free_deviceid_node(struct nfs4_deviceid_node *d)
1092 {
1093 	nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1094 }
1095 
1096 static struct pnfs_layout_hdr *
1097 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1098 {
1099 	struct nfs4_filelayout *flo;
1100 
1101 	flo = kzalloc(sizeof(*flo), gfp_flags);
1102 	return flo != NULL ? &flo->generic_hdr : NULL;
1103 }
1104 
1105 static void
1106 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1107 {
1108 	kfree(FILELAYOUT_FROM_HDR(lo));
1109 }
1110 
1111 static struct pnfs_ds_commit_info *
1112 filelayout_get_ds_info(struct inode *inode)
1113 {
1114 	struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1115 
1116 	if (layout == NULL)
1117 		return NULL;
1118 	else
1119 		return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1120 }
1121 
1122 static struct pnfs_layoutdriver_type filelayout_type = {
1123 	.id			= LAYOUT_NFSV4_1_FILES,
1124 	.name			= "LAYOUT_NFSV4_1_FILES",
1125 	.owner			= THIS_MODULE,
1126 	.alloc_layout_hdr	= filelayout_alloc_layout_hdr,
1127 	.free_layout_hdr	= filelayout_free_layout_hdr,
1128 	.alloc_lseg		= filelayout_alloc_lseg,
1129 	.free_lseg		= filelayout_free_lseg,
1130 	.pg_read_ops		= &filelayout_pg_read_ops,
1131 	.pg_write_ops		= &filelayout_pg_write_ops,
1132 	.get_ds_info		= &filelayout_get_ds_info,
1133 	.mark_request_commit	= filelayout_mark_request_commit,
1134 	.clear_request_commit	= pnfs_generic_clear_request_commit,
1135 	.scan_commit_lists	= pnfs_generic_scan_commit_lists,
1136 	.recover_commit_reqs	= pnfs_generic_recover_commit_reqs,
1137 	.search_commit_reqs	= filelayout_search_commit_reqs,
1138 	.commit_pagelist	= filelayout_commit_pagelist,
1139 	.read_pagelist		= filelayout_read_pagelist,
1140 	.write_pagelist		= filelayout_write_pagelist,
1141 	.alloc_deviceid_node	= filelayout_alloc_deviceid_node,
1142 	.free_deviceid_node	= filelayout_free_deviceid_node,
1143 	.sync			= pnfs_nfs_generic_sync,
1144 };
1145 
1146 static int __init nfs4filelayout_init(void)
1147 {
1148 	printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1149 	       __func__);
1150 	return pnfs_register_layoutdriver(&filelayout_type);
1151 }
1152 
1153 static void __exit nfs4filelayout_exit(void)
1154 {
1155 	printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1156 	       __func__);
1157 	pnfs_unregister_layoutdriver(&filelayout_type);
1158 }
1159 
1160 MODULE_ALIAS("nfs-layouttype4-1");
1161 
1162 module_init(nfs4filelayout_init);
1163 module_exit(nfs4filelayout_exit);
1164