xref: /openbmc/linux/fs/nfs/filelayout/filelayout.c (revision adb57164)
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 static const struct pnfs_commit_ops filelayout_commit_ops;
53 
54 static loff_t
55 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
56 			    loff_t offset)
57 {
58 	u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
59 	u64 stripe_no;
60 	u32 rem;
61 
62 	offset -= flseg->pattern_offset;
63 	stripe_no = div_u64(offset, stripe_width);
64 	div_u64_rem(offset, flseg->stripe_unit, &rem);
65 
66 	return stripe_no * flseg->stripe_unit + rem;
67 }
68 
69 /* This function is used by the layout driver to calculate the
70  * offset of the file on the dserver based on whether the
71  * layout type is STRIPE_DENSE or STRIPE_SPARSE
72  */
73 static loff_t
74 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
75 {
76 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
77 
78 	switch (flseg->stripe_type) {
79 	case STRIPE_SPARSE:
80 		return offset;
81 
82 	case STRIPE_DENSE:
83 		return filelayout_get_dense_offset(flseg, offset);
84 	}
85 
86 	BUG();
87 }
88 
89 static void filelayout_reset_write(struct nfs_pgio_header *hdr)
90 {
91 	struct rpc_task *task = &hdr->task;
92 
93 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
94 		dprintk("%s Reset task %5u for i/o through MDS "
95 			"(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
96 			hdr->task.tk_pid,
97 			hdr->inode->i_sb->s_id,
98 			(unsigned long long)NFS_FILEID(hdr->inode),
99 			hdr->args.count,
100 			(unsigned long long)hdr->args.offset);
101 
102 		task->tk_status = pnfs_write_done_resend_to_mds(hdr);
103 	}
104 }
105 
106 static void filelayout_reset_read(struct nfs_pgio_header *hdr)
107 {
108 	struct rpc_task *task = &hdr->task;
109 
110 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
111 		dprintk("%s Reset task %5u for i/o through MDS "
112 			"(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
113 			hdr->task.tk_pid,
114 			hdr->inode->i_sb->s_id,
115 			(unsigned long long)NFS_FILEID(hdr->inode),
116 			hdr->args.count,
117 			(unsigned long long)hdr->args.offset);
118 
119 		task->tk_status = pnfs_read_done_resend_to_mds(hdr);
120 	}
121 }
122 
123 static int filelayout_async_handle_error(struct rpc_task *task,
124 					 struct nfs4_state *state,
125 					 struct nfs_client *clp,
126 					 struct pnfs_layout_segment *lseg)
127 {
128 	struct pnfs_layout_hdr *lo = lseg->pls_layout;
129 	struct inode *inode = lo->plh_inode;
130 	struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
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 	/* DS session errors */
138 	case -NFS4ERR_BADSESSION:
139 	case -NFS4ERR_BADSLOT:
140 	case -NFS4ERR_BAD_HIGH_SLOT:
141 	case -NFS4ERR_DEADSESSION:
142 	case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
143 	case -NFS4ERR_SEQ_FALSE_RETRY:
144 	case -NFS4ERR_SEQ_MISORDERED:
145 		dprintk("%s ERROR %d, Reset session. Exchangeid "
146 			"flags 0x%x\n", __func__, task->tk_status,
147 			clp->cl_exchange_flags);
148 		nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
149 		break;
150 	case -NFS4ERR_DELAY:
151 	case -NFS4ERR_GRACE:
152 		rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
153 		break;
154 	case -NFS4ERR_RETRY_UNCACHED_REP:
155 		break;
156 	/* Invalidate Layout errors */
157 	case -NFS4ERR_ACCESS:
158 	case -NFS4ERR_PNFS_NO_LAYOUT:
159 	case -ESTALE:           /* mapped NFS4ERR_STALE */
160 	case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
161 	case -EISDIR:           /* mapped NFS4ERR_ISDIR */
162 	case -NFS4ERR_FHEXPIRED:
163 	case -NFS4ERR_WRONG_TYPE:
164 		dprintk("%s Invalid layout error %d\n", __func__,
165 			task->tk_status);
166 		/*
167 		 * Destroy layout so new i/o will get a new layout.
168 		 * Layout will not be destroyed until all current lseg
169 		 * references are put. Mark layout as invalid to resend failed
170 		 * i/o and all i/o waiting on the slot table to the MDS until
171 		 * layout is destroyed and a new valid layout is obtained.
172 		 */
173 		pnfs_destroy_layout(NFS_I(inode));
174 		rpc_wake_up(&tbl->slot_tbl_waitq);
175 		goto reset;
176 	/* RPC connection errors */
177 	case -ECONNREFUSED:
178 	case -EHOSTDOWN:
179 	case -EHOSTUNREACH:
180 	case -ENETUNREACH:
181 	case -EIO:
182 	case -ETIMEDOUT:
183 	case -EPIPE:
184 		dprintk("%s DS connection error %d\n", __func__,
185 			task->tk_status);
186 		nfs4_mark_deviceid_unavailable(devid);
187 		pnfs_error_mark_layout_for_return(inode, lseg);
188 		pnfs_set_lo_fail(lseg);
189 		rpc_wake_up(&tbl->slot_tbl_waitq);
190 		/* fall through */
191 	default:
192 reset:
193 		dprintk("%s Retry through MDS. Error %d\n", __func__,
194 			task->tk_status);
195 		return -NFS4ERR_RESET_TO_MDS;
196 	}
197 	task->tk_status = 0;
198 	return -EAGAIN;
199 }
200 
201 /* NFS_PROTO call done callback routines */
202 
203 static int filelayout_read_done_cb(struct rpc_task *task,
204 				struct nfs_pgio_header *hdr)
205 {
206 	int err;
207 
208 	trace_nfs4_pnfs_read(hdr, task->tk_status);
209 	err = filelayout_async_handle_error(task, hdr->args.context->state,
210 					    hdr->ds_clp, hdr->lseg);
211 
212 	switch (err) {
213 	case -NFS4ERR_RESET_TO_MDS:
214 		filelayout_reset_read(hdr);
215 		return task->tk_status;
216 	case -EAGAIN:
217 		rpc_restart_call_prepare(task);
218 		return -EAGAIN;
219 	}
220 
221 	return 0;
222 }
223 
224 /*
225  * We reference the rpc_cred of the first WRITE that triggers the need for
226  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
227  * rfc5661 is not clear about which credential should be used.
228  */
229 static void
230 filelayout_set_layoutcommit(struct nfs_pgio_header *hdr)
231 {
232 	loff_t end_offs = 0;
233 
234 	if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
235 	    hdr->res.verf->committed == NFS_FILE_SYNC)
236 		return;
237 	if (hdr->res.verf->committed == NFS_DATA_SYNC)
238 		end_offs = hdr->mds_offset + (loff_t)hdr->res.count;
239 
240 	/* Note: if the write is unstable, don't set end_offs until commit */
241 	pnfs_set_layoutcommit(hdr->inode, hdr->lseg, end_offs);
242 	dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
243 		(unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
244 }
245 
246 bool
247 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
248 {
249 	return filelayout_test_devid_invalid(node) ||
250 		nfs4_test_deviceid_unavailable(node);
251 }
252 
253 static bool
254 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
255 {
256 	struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
257 
258 	return filelayout_test_devid_unavailable(node);
259 }
260 
261 /*
262  * Call ops for the async read/write cases
263  * In the case of dense layouts, the offset needs to be reset to its
264  * original value.
265  */
266 static void filelayout_read_prepare(struct rpc_task *task, void *data)
267 {
268 	struct nfs_pgio_header *hdr = data;
269 
270 	if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
271 		rpc_exit(task, -EIO);
272 		return;
273 	}
274 	if (filelayout_reset_to_mds(hdr->lseg)) {
275 		dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
276 		filelayout_reset_read(hdr);
277 		rpc_exit(task, 0);
278 		return;
279 	}
280 	hdr->pgio_done_cb = filelayout_read_done_cb;
281 
282 	if (nfs4_setup_sequence(hdr->ds_clp,
283 			&hdr->args.seq_args,
284 			&hdr->res.seq_res,
285 			task))
286 		return;
287 	if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
288 			hdr->args.lock_context, FMODE_READ) == -EIO)
289 		rpc_exit(task, -EIO); /* lost lock, terminate I/O */
290 }
291 
292 static void filelayout_read_call_done(struct rpc_task *task, void *data)
293 {
294 	struct nfs_pgio_header *hdr = data;
295 
296 	dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
297 
298 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
299 	    task->tk_status == 0) {
300 		nfs41_sequence_done(task, &hdr->res.seq_res);
301 		return;
302 	}
303 
304 	/* Note this may cause RPC to be resent */
305 	hdr->mds_ops->rpc_call_done(task, data);
306 }
307 
308 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
309 {
310 	struct nfs_pgio_header *hdr = data;
311 
312 	rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
313 }
314 
315 static int filelayout_write_done_cb(struct rpc_task *task,
316 				struct nfs_pgio_header *hdr)
317 {
318 	int err;
319 
320 	trace_nfs4_pnfs_write(hdr, task->tk_status);
321 	err = filelayout_async_handle_error(task, hdr->args.context->state,
322 					    hdr->ds_clp, hdr->lseg);
323 
324 	switch (err) {
325 	case -NFS4ERR_RESET_TO_MDS:
326 		filelayout_reset_write(hdr);
327 		return task->tk_status;
328 	case -EAGAIN:
329 		rpc_restart_call_prepare(task);
330 		return -EAGAIN;
331 	}
332 
333 	filelayout_set_layoutcommit(hdr);
334 
335 	/* zero out the fattr */
336 	hdr->fattr.valid = 0;
337 	if (task->tk_status >= 0)
338 		nfs_writeback_update_inode(hdr);
339 
340 	return 0;
341 }
342 
343 static int filelayout_commit_done_cb(struct rpc_task *task,
344 				     struct nfs_commit_data *data)
345 {
346 	int err;
347 
348 	trace_nfs4_pnfs_commit_ds(data, task->tk_status);
349 	err = filelayout_async_handle_error(task, NULL, data->ds_clp,
350 					    data->lseg);
351 
352 	switch (err) {
353 	case -NFS4ERR_RESET_TO_MDS:
354 		pnfs_generic_prepare_to_resend_writes(data);
355 		return -EAGAIN;
356 	case -EAGAIN:
357 		rpc_restart_call_prepare(task);
358 		return -EAGAIN;
359 	}
360 
361 	pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
362 
363 	return 0;
364 }
365 
366 static void filelayout_write_prepare(struct rpc_task *task, void *data)
367 {
368 	struct nfs_pgio_header *hdr = data;
369 
370 	if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
371 		rpc_exit(task, -EIO);
372 		return;
373 	}
374 	if (filelayout_reset_to_mds(hdr->lseg)) {
375 		dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
376 		filelayout_reset_write(hdr);
377 		rpc_exit(task, 0);
378 		return;
379 	}
380 	if (nfs4_setup_sequence(hdr->ds_clp,
381 			&hdr->args.seq_args,
382 			&hdr->res.seq_res,
383 			task))
384 		return;
385 	if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
386 			hdr->args.lock_context, FMODE_WRITE) == -EIO)
387 		rpc_exit(task, -EIO); /* lost lock, terminate I/O */
388 }
389 
390 static void filelayout_write_call_done(struct rpc_task *task, void *data)
391 {
392 	struct nfs_pgio_header *hdr = data;
393 
394 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
395 	    task->tk_status == 0) {
396 		nfs41_sequence_done(task, &hdr->res.seq_res);
397 		return;
398 	}
399 
400 	/* Note this may cause RPC to be resent */
401 	hdr->mds_ops->rpc_call_done(task, data);
402 }
403 
404 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
405 {
406 	struct nfs_pgio_header *hdr = data;
407 
408 	rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
409 }
410 
411 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
412 {
413 	struct nfs_commit_data *wdata = data;
414 
415 	nfs4_setup_sequence(wdata->ds_clp,
416 			&wdata->args.seq_args,
417 			&wdata->res.seq_res,
418 			task);
419 }
420 
421 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
422 {
423 	struct nfs_commit_data *cdata = data;
424 
425 	rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
426 }
427 
428 static const struct rpc_call_ops filelayout_read_call_ops = {
429 	.rpc_call_prepare = filelayout_read_prepare,
430 	.rpc_call_done = filelayout_read_call_done,
431 	.rpc_count_stats = filelayout_read_count_stats,
432 	.rpc_release = pnfs_generic_rw_release,
433 };
434 
435 static const struct rpc_call_ops filelayout_write_call_ops = {
436 	.rpc_call_prepare = filelayout_write_prepare,
437 	.rpc_call_done = filelayout_write_call_done,
438 	.rpc_count_stats = filelayout_write_count_stats,
439 	.rpc_release = pnfs_generic_rw_release,
440 };
441 
442 static const struct rpc_call_ops filelayout_commit_call_ops = {
443 	.rpc_call_prepare = filelayout_commit_prepare,
444 	.rpc_call_done = pnfs_generic_write_commit_done,
445 	.rpc_count_stats = filelayout_commit_count_stats,
446 	.rpc_release = pnfs_generic_commit_release,
447 };
448 
449 static enum pnfs_try_status
450 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
451 {
452 	struct pnfs_layout_segment *lseg = hdr->lseg;
453 	struct nfs4_pnfs_ds *ds;
454 	struct rpc_clnt *ds_clnt;
455 	loff_t offset = hdr->args.offset;
456 	u32 j, idx;
457 	struct nfs_fh *fh;
458 
459 	dprintk("--> %s ino %lu pgbase %u req %zu@%llu\n",
460 		__func__, hdr->inode->i_ino,
461 		hdr->args.pgbase, (size_t)hdr->args.count, offset);
462 
463 	/* Retrieve the correct rpc_client for the byte range */
464 	j = nfs4_fl_calc_j_index(lseg, offset);
465 	idx = nfs4_fl_calc_ds_index(lseg, j);
466 	ds = nfs4_fl_prepare_ds(lseg, idx);
467 	if (!ds)
468 		return PNFS_NOT_ATTEMPTED;
469 
470 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
471 	if (IS_ERR(ds_clnt))
472 		return PNFS_NOT_ATTEMPTED;
473 
474 	dprintk("%s USE DS: %s cl_count %d\n", __func__,
475 		ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count));
476 
477 	/* No multipath support. Use first DS */
478 	refcount_inc(&ds->ds_clp->cl_count);
479 	hdr->ds_clp = ds->ds_clp;
480 	hdr->ds_commit_idx = idx;
481 	fh = nfs4_fl_select_ds_fh(lseg, j);
482 	if (fh)
483 		hdr->args.fh = fh;
484 
485 	hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
486 	hdr->mds_offset = offset;
487 
488 	/* Perform an asynchronous read to ds */
489 	nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
490 			  NFS_PROTO(hdr->inode), &filelayout_read_call_ops,
491 			  0, RPC_TASK_SOFTCONN);
492 	return PNFS_ATTEMPTED;
493 }
494 
495 /* Perform async writes. */
496 static enum pnfs_try_status
497 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
498 {
499 	struct pnfs_layout_segment *lseg = hdr->lseg;
500 	struct nfs4_pnfs_ds *ds;
501 	struct rpc_clnt *ds_clnt;
502 	loff_t offset = hdr->args.offset;
503 	u32 j, idx;
504 	struct nfs_fh *fh;
505 
506 	/* Retrieve the correct rpc_client for the byte range */
507 	j = nfs4_fl_calc_j_index(lseg, offset);
508 	idx = nfs4_fl_calc_ds_index(lseg, j);
509 	ds = nfs4_fl_prepare_ds(lseg, idx);
510 	if (!ds)
511 		return PNFS_NOT_ATTEMPTED;
512 
513 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
514 	if (IS_ERR(ds_clnt))
515 		return PNFS_NOT_ATTEMPTED;
516 
517 	dprintk("%s ino %lu sync %d req %zu@%llu DS: %s cl_count %d\n",
518 		__func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
519 		offset, ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count));
520 
521 	hdr->pgio_done_cb = filelayout_write_done_cb;
522 	refcount_inc(&ds->ds_clp->cl_count);
523 	hdr->ds_clp = ds->ds_clp;
524 	hdr->ds_commit_idx = idx;
525 	fh = nfs4_fl_select_ds_fh(lseg, j);
526 	if (fh)
527 		hdr->args.fh = fh;
528 	hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
529 
530 	/* Perform an asynchronous write */
531 	nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
532 			  NFS_PROTO(hdr->inode), &filelayout_write_call_ops,
533 			  sync, RPC_TASK_SOFTCONN);
534 	return PNFS_ATTEMPTED;
535 }
536 
537 static int
538 filelayout_check_deviceid(struct pnfs_layout_hdr *lo,
539 			  struct nfs4_filelayout_segment *fl,
540 			  gfp_t gfp_flags)
541 {
542 	struct nfs4_deviceid_node *d;
543 	struct nfs4_file_layout_dsaddr *dsaddr;
544 	int status = -EINVAL;
545 
546 	/* Is the deviceid already set? If so, we're good. */
547 	if (fl->dsaddr != NULL)
548 		return 0;
549 
550 	/* find and reference the deviceid */
551 	d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), &fl->deviceid,
552 			lo->plh_lc_cred, gfp_flags);
553 	if (d == NULL)
554 		goto out;
555 
556 	dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
557 	/* Found deviceid is unavailable */
558 	if (filelayout_test_devid_unavailable(&dsaddr->id_node))
559 		goto out_put;
560 
561 	if (fl->first_stripe_index >= dsaddr->stripe_count) {
562 		dprintk("%s Bad first_stripe_index %u\n",
563 				__func__, fl->first_stripe_index);
564 		goto out_put;
565 	}
566 
567 	if ((fl->stripe_type == STRIPE_SPARSE &&
568 	    fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
569 	    (fl->stripe_type == STRIPE_DENSE &&
570 	    fl->num_fh != dsaddr->stripe_count)) {
571 		dprintk("%s num_fh %u not valid for given packing\n",
572 			__func__, fl->num_fh);
573 		goto out_put;
574 	}
575 	status = 0;
576 
577 	/*
578 	 * Atomic compare and xchange to ensure we don't scribble
579 	 * over a non-NULL pointer.
580 	 */
581 	if (cmpxchg(&fl->dsaddr, NULL, dsaddr) != NULL)
582 		goto out_put;
583 out:
584 	return status;
585 out_put:
586 	nfs4_fl_put_deviceid(dsaddr);
587 	goto out;
588 }
589 
590 /*
591  * filelayout_check_layout()
592  *
593  * Make sure layout segment parameters are sane WRT the device.
594  * At this point no generic layer initialization of the lseg has occurred,
595  * and nothing has been added to the layout_hdr cache.
596  *
597  */
598 static int
599 filelayout_check_layout(struct pnfs_layout_hdr *lo,
600 			struct nfs4_filelayout_segment *fl,
601 			struct nfs4_layoutget_res *lgr,
602 			gfp_t gfp_flags)
603 {
604 	int status = -EINVAL;
605 
606 	dprintk("--> %s\n", __func__);
607 
608 	/* FIXME: remove this check when layout segment support is added */
609 	if (lgr->range.offset != 0 ||
610 	    lgr->range.length != NFS4_MAX_UINT64) {
611 		dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
612 			__func__);
613 		goto out;
614 	}
615 
616 	if (fl->pattern_offset > lgr->range.offset) {
617 		dprintk("%s pattern_offset %lld too large\n",
618 				__func__, fl->pattern_offset);
619 		goto out;
620 	}
621 
622 	if (!fl->stripe_unit) {
623 		dprintk("%s Invalid stripe unit (%u)\n",
624 			__func__, fl->stripe_unit);
625 		goto out;
626 	}
627 
628 	status = 0;
629 out:
630 	dprintk("--> %s returns %d\n", __func__, status);
631 	return status;
632 }
633 
634 static void _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
635 {
636 	int i;
637 
638 	if (fl->fh_array) {
639 		for (i = 0; i < fl->num_fh; i++) {
640 			if (!fl->fh_array[i])
641 				break;
642 			kfree(fl->fh_array[i]);
643 		}
644 		kfree(fl->fh_array);
645 	}
646 	kfree(fl);
647 }
648 
649 static int
650 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
651 			 struct nfs4_filelayout_segment *fl,
652 			 struct nfs4_layoutget_res *lgr,
653 			 gfp_t gfp_flags)
654 {
655 	struct xdr_stream stream;
656 	struct xdr_buf buf;
657 	struct page *scratch;
658 	__be32 *p;
659 	uint32_t nfl_util;
660 	int i;
661 
662 	dprintk("%s: set_layout_map Begin\n", __func__);
663 
664 	scratch = alloc_page(gfp_flags);
665 	if (!scratch)
666 		return -ENOMEM;
667 
668 	xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
669 	xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
670 
671 	/* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
672 	 * num_fh (4) */
673 	p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
674 	if (unlikely(!p))
675 		goto out_err;
676 
677 	memcpy(&fl->deviceid, p, sizeof(fl->deviceid));
678 	p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
679 	nfs4_print_deviceid(&fl->deviceid);
680 
681 	nfl_util = be32_to_cpup(p++);
682 	if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
683 		fl->commit_through_mds = 1;
684 	if (nfl_util & NFL4_UFLG_DENSE)
685 		fl->stripe_type = STRIPE_DENSE;
686 	else
687 		fl->stripe_type = STRIPE_SPARSE;
688 	fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
689 
690 	fl->first_stripe_index = be32_to_cpup(p++);
691 	p = xdr_decode_hyper(p, &fl->pattern_offset);
692 	fl->num_fh = be32_to_cpup(p++);
693 
694 	dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
695 		__func__, nfl_util, fl->num_fh, fl->first_stripe_index,
696 		fl->pattern_offset);
697 
698 	/* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
699 	 * Futher checking is done in filelayout_check_layout */
700 	if (fl->num_fh >
701 	    max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
702 		goto out_err;
703 
704 	if (fl->num_fh > 0) {
705 		fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
706 				       gfp_flags);
707 		if (!fl->fh_array)
708 			goto out_err;
709 	}
710 
711 	for (i = 0; i < fl->num_fh; i++) {
712 		/* Do we want to use a mempool here? */
713 		fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
714 		if (!fl->fh_array[i])
715 			goto out_err;
716 
717 		p = xdr_inline_decode(&stream, 4);
718 		if (unlikely(!p))
719 			goto out_err;
720 		fl->fh_array[i]->size = be32_to_cpup(p++);
721 		if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
722 			printk(KERN_ERR "NFS: Too big fh %d received %d\n",
723 			       i, fl->fh_array[i]->size);
724 			goto out_err;
725 		}
726 
727 		p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
728 		if (unlikely(!p))
729 			goto out_err;
730 		memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
731 		dprintk("DEBUG: %s: fh len %d\n", __func__,
732 			fl->fh_array[i]->size);
733 	}
734 
735 	__free_page(scratch);
736 	return 0;
737 
738 out_err:
739 	__free_page(scratch);
740 	return -EIO;
741 }
742 
743 static void
744 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
745 {
746 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
747 
748 	dprintk("--> %s\n", __func__);
749 	if (fl->dsaddr != NULL)
750 		nfs4_fl_put_deviceid(fl->dsaddr);
751 	/* This assumes a single RW lseg */
752 	if (lseg->pls_range.iomode == IOMODE_RW) {
753 		struct nfs4_filelayout *flo;
754 		struct inode *inode;
755 
756 		flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
757 		inode = flo->generic_hdr.plh_inode;
758 		spin_lock(&inode->i_lock);
759 		pnfs_generic_ds_cinfo_release_lseg(&flo->commit_info, lseg);
760 		spin_unlock(&inode->i_lock);
761 	}
762 	_filelayout_free_lseg(fl);
763 }
764 
765 static struct pnfs_layout_segment *
766 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
767 		      struct nfs4_layoutget_res *lgr,
768 		      gfp_t gfp_flags)
769 {
770 	struct nfs4_filelayout_segment *fl;
771 	int rc;
772 
773 	dprintk("--> %s\n", __func__);
774 	fl = kzalloc(sizeof(*fl), gfp_flags);
775 	if (!fl)
776 		return NULL;
777 
778 	rc = filelayout_decode_layout(layoutid, fl, lgr, gfp_flags);
779 	if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, gfp_flags)) {
780 		_filelayout_free_lseg(fl);
781 		return NULL;
782 	}
783 	return &fl->generic_hdr;
784 }
785 
786 /*
787  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
788  *
789  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
790  * of bytes (maximum @req->wb_bytes) that can be coalesced.
791  */
792 static size_t
793 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
794 		   struct nfs_page *req)
795 {
796 	unsigned int size;
797 	u64 p_stripe, r_stripe;
798 	u32 stripe_offset;
799 	u64 segment_offset = pgio->pg_lseg->pls_range.offset;
800 	u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
801 
802 	/* calls nfs_generic_pg_test */
803 	size = pnfs_generic_pg_test(pgio, prev, req);
804 	if (!size)
805 		return 0;
806 
807 	/* see if req and prev are in the same stripe */
808 	if (prev) {
809 		p_stripe = (u64)req_offset(prev) - segment_offset;
810 		r_stripe = (u64)req_offset(req) - segment_offset;
811 		do_div(p_stripe, stripe_unit);
812 		do_div(r_stripe, stripe_unit);
813 
814 		if (p_stripe != r_stripe)
815 			return 0;
816 	}
817 
818 	/* calculate remaining bytes in the current stripe */
819 	div_u64_rem((u64)req_offset(req) - segment_offset,
820 			stripe_unit,
821 			&stripe_offset);
822 	WARN_ON_ONCE(stripe_offset > stripe_unit);
823 	if (stripe_offset >= stripe_unit)
824 		return 0;
825 	return min(stripe_unit - (unsigned int)stripe_offset, size);
826 }
827 
828 static struct pnfs_layout_segment *
829 fl_pnfs_update_layout(struct inode *ino,
830 		      struct nfs_open_context *ctx,
831 		      loff_t pos,
832 		      u64 count,
833 		      enum pnfs_iomode iomode,
834 		      bool strict_iomode,
835 		      gfp_t gfp_flags)
836 {
837 	struct pnfs_layout_segment *lseg = NULL;
838 	struct pnfs_layout_hdr *lo;
839 	struct nfs4_filelayout_segment *fl;
840 	int status;
841 
842 	lseg = pnfs_update_layout(ino, ctx, pos, count, iomode, strict_iomode,
843 				  gfp_flags);
844 	if (IS_ERR_OR_NULL(lseg))
845 		goto out;
846 
847 	lo = NFS_I(ino)->layout;
848 	fl = FILELAYOUT_LSEG(lseg);
849 
850 	status = filelayout_check_deviceid(lo, fl, gfp_flags);
851 	if (status) {
852 		pnfs_put_lseg(lseg);
853 		lseg = NULL;
854 	}
855 out:
856 	return lseg;
857 }
858 
859 static void
860 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
861 			struct nfs_page *req)
862 {
863 	pnfs_generic_pg_check_layout(pgio);
864 	if (!pgio->pg_lseg) {
865 		pgio->pg_lseg = fl_pnfs_update_layout(pgio->pg_inode,
866 						      nfs_req_openctx(req),
867 						      0,
868 						      NFS4_MAX_UINT64,
869 						      IOMODE_READ,
870 						      false,
871 						      GFP_KERNEL);
872 		if (IS_ERR(pgio->pg_lseg)) {
873 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
874 			pgio->pg_lseg = NULL;
875 			return;
876 		}
877 	}
878 	/* If no lseg, fall back to read through mds */
879 	if (pgio->pg_lseg == NULL)
880 		nfs_pageio_reset_read_mds(pgio);
881 }
882 
883 static void
884 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
885 			 struct nfs_page *req)
886 {
887 	pnfs_generic_pg_check_layout(pgio);
888 	if (!pgio->pg_lseg) {
889 		pgio->pg_lseg = fl_pnfs_update_layout(pgio->pg_inode,
890 						      nfs_req_openctx(req),
891 						      0,
892 						      NFS4_MAX_UINT64,
893 						      IOMODE_RW,
894 						      false,
895 						      GFP_NOFS);
896 		if (IS_ERR(pgio->pg_lseg)) {
897 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
898 			pgio->pg_lseg = NULL;
899 			return;
900 		}
901 	}
902 
903 	/* If no lseg, fall back to write through mds */
904 	if (pgio->pg_lseg == NULL)
905 		nfs_pageio_reset_write_mds(pgio);
906 }
907 
908 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
909 	.pg_init = filelayout_pg_init_read,
910 	.pg_test = filelayout_pg_test,
911 	.pg_doio = pnfs_generic_pg_readpages,
912 	.pg_cleanup = pnfs_generic_pg_cleanup,
913 };
914 
915 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
916 	.pg_init = filelayout_pg_init_write,
917 	.pg_test = filelayout_pg_test,
918 	.pg_doio = pnfs_generic_pg_writepages,
919 	.pg_cleanup = pnfs_generic_pg_cleanup,
920 };
921 
922 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
923 {
924 	if (fl->stripe_type == STRIPE_SPARSE)
925 		return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
926 	else
927 		return j;
928 }
929 
930 static void
931 filelayout_mark_request_commit(struct nfs_page *req,
932 			       struct pnfs_layout_segment *lseg,
933 			       struct nfs_commit_info *cinfo,
934 			       u32 ds_commit_idx)
935 
936 {
937 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
938 	u32 i, j;
939 
940 	if (fl->commit_through_mds) {
941 		nfs_request_add_commit_list(req, cinfo);
942 	} else {
943 		/* Note that we are calling nfs4_fl_calc_j_index on each page
944 		 * that ends up being committed to a data server.  An attractive
945 		 * alternative is to add a field to nfs_write_data and nfs_page
946 		 * to store the value calculated in filelayout_write_pagelist
947 		 * and just use that here.
948 		 */
949 		j = nfs4_fl_calc_j_index(lseg, req_offset(req));
950 		i = select_bucket_index(fl, j);
951 		pnfs_layout_mark_request_commit(req, lseg, cinfo, i);
952 	}
953 }
954 
955 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
956 {
957 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
958 
959 	if (flseg->stripe_type == STRIPE_SPARSE)
960 		return i;
961 	else
962 		return nfs4_fl_calc_ds_index(lseg, i);
963 }
964 
965 static struct nfs_fh *
966 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
967 {
968 	struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
969 
970 	if (flseg->stripe_type == STRIPE_SPARSE) {
971 		if (flseg->num_fh == 1)
972 			i = 0;
973 		else if (flseg->num_fh == 0)
974 			/* Use the MDS OPEN fh set in nfs_read_rpcsetup */
975 			return NULL;
976 	}
977 	return flseg->fh_array[i];
978 }
979 
980 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
981 {
982 	struct pnfs_layout_segment *lseg = data->lseg;
983 	struct nfs4_pnfs_ds *ds;
984 	struct rpc_clnt *ds_clnt;
985 	u32 idx;
986 	struct nfs_fh *fh;
987 
988 	idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
989 	ds = nfs4_fl_prepare_ds(lseg, idx);
990 	if (!ds)
991 		goto out_err;
992 
993 	ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
994 	if (IS_ERR(ds_clnt))
995 		goto out_err;
996 
997 	dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
998 		data->inode->i_ino, how, refcount_read(&ds->ds_clp->cl_count));
999 	data->commit_done_cb = filelayout_commit_done_cb;
1000 	refcount_inc(&ds->ds_clp->cl_count);
1001 	data->ds_clp = ds->ds_clp;
1002 	fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1003 	if (fh)
1004 		data->args.fh = fh;
1005 	return nfs_initiate_commit(ds_clnt, data, NFS_PROTO(data->inode),
1006 				   &filelayout_commit_call_ops, how,
1007 				   RPC_TASK_SOFTCONN);
1008 out_err:
1009 	pnfs_generic_prepare_to_resend_writes(data);
1010 	pnfs_generic_commit_release(data);
1011 	return -EAGAIN;
1012 }
1013 
1014 static int
1015 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1016 			   int how, struct nfs_commit_info *cinfo)
1017 {
1018 	return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1019 					    filelayout_initiate_commit);
1020 }
1021 
1022 static struct nfs4_deviceid_node *
1023 filelayout_alloc_deviceid_node(struct nfs_server *server,
1024 		struct pnfs_device *pdev, gfp_t gfp_flags)
1025 {
1026 	struct nfs4_file_layout_dsaddr *dsaddr;
1027 
1028 	dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1029 	if (!dsaddr)
1030 		return NULL;
1031 	return &dsaddr->id_node;
1032 }
1033 
1034 static void
1035 filelayout_free_deviceid_node(struct nfs4_deviceid_node *d)
1036 {
1037 	nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1038 }
1039 
1040 static struct pnfs_layout_hdr *
1041 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1042 {
1043 	struct nfs4_filelayout *flo;
1044 
1045 	flo = kzalloc(sizeof(*flo), gfp_flags);
1046 	if (flo == NULL)
1047 		return NULL;
1048 	pnfs_init_ds_commit_info(&flo->commit_info);
1049 	flo->commit_info.ops = &filelayout_commit_ops;
1050 	return &flo->generic_hdr;
1051 }
1052 
1053 static void
1054 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1055 {
1056 	kfree_rcu(FILELAYOUT_FROM_HDR(lo), generic_hdr.plh_rcu);
1057 }
1058 
1059 static struct pnfs_ds_commit_info *
1060 filelayout_get_ds_info(struct inode *inode)
1061 {
1062 	struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1063 
1064 	if (layout == NULL)
1065 		return NULL;
1066 	else
1067 		return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1068 }
1069 
1070 static void
1071 filelayout_setup_ds_info(struct pnfs_ds_commit_info *fl_cinfo,
1072 		struct pnfs_layout_segment *lseg)
1073 {
1074 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1075 	struct inode *inode = lseg->pls_layout->plh_inode;
1076 	struct pnfs_commit_array *array, *new;
1077 	unsigned int size = (fl->stripe_type == STRIPE_SPARSE) ?
1078 		fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
1079 
1080 	new = pnfs_alloc_commit_array(size, GFP_NOIO);
1081 	if (new) {
1082 		spin_lock(&inode->i_lock);
1083 		array = pnfs_add_commit_array(fl_cinfo, new, lseg);
1084 		spin_unlock(&inode->i_lock);
1085 		if (array != new)
1086 			pnfs_free_commit_array(new);
1087 	}
1088 }
1089 
1090 static void
1091 filelayout_release_ds_info(struct pnfs_ds_commit_info *fl_cinfo,
1092 		struct inode *inode)
1093 {
1094 	spin_lock(&inode->i_lock);
1095 	pnfs_generic_ds_cinfo_destroy(fl_cinfo);
1096 	spin_unlock(&inode->i_lock);
1097 }
1098 
1099 static const struct pnfs_commit_ops filelayout_commit_ops = {
1100 	.setup_ds_info		= filelayout_setup_ds_info,
1101 	.release_ds_info	= filelayout_release_ds_info,
1102 	.mark_request_commit	= filelayout_mark_request_commit,
1103 	.clear_request_commit	= pnfs_generic_clear_request_commit,
1104 	.scan_commit_lists	= pnfs_generic_scan_commit_lists,
1105 	.recover_commit_reqs	= pnfs_generic_recover_commit_reqs,
1106 	.search_commit_reqs	= pnfs_generic_search_commit_reqs,
1107 	.commit_pagelist	= filelayout_commit_pagelist,
1108 };
1109 
1110 static struct pnfs_layoutdriver_type filelayout_type = {
1111 	.id			= LAYOUT_NFSV4_1_FILES,
1112 	.name			= "LAYOUT_NFSV4_1_FILES",
1113 	.owner			= THIS_MODULE,
1114 	.flags			= PNFS_LAYOUTGET_ON_OPEN,
1115 	.max_layoutget_response	= 4096, /* 1 page or so... */
1116 	.alloc_layout_hdr	= filelayout_alloc_layout_hdr,
1117 	.free_layout_hdr	= filelayout_free_layout_hdr,
1118 	.alloc_lseg		= filelayout_alloc_lseg,
1119 	.free_lseg		= filelayout_free_lseg,
1120 	.pg_read_ops		= &filelayout_pg_read_ops,
1121 	.pg_write_ops		= &filelayout_pg_write_ops,
1122 	.get_ds_info		= &filelayout_get_ds_info,
1123 	.read_pagelist		= filelayout_read_pagelist,
1124 	.write_pagelist		= filelayout_write_pagelist,
1125 	.alloc_deviceid_node	= filelayout_alloc_deviceid_node,
1126 	.free_deviceid_node	= filelayout_free_deviceid_node,
1127 	.sync			= pnfs_nfs_generic_sync,
1128 };
1129 
1130 static int __init nfs4filelayout_init(void)
1131 {
1132 	printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1133 	       __func__);
1134 	return pnfs_register_layoutdriver(&filelayout_type);
1135 }
1136 
1137 static void __exit nfs4filelayout_exit(void)
1138 {
1139 	printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1140 	       __func__);
1141 	pnfs_unregister_layoutdriver(&filelayout_type);
1142 }
1143 
1144 MODULE_ALIAS("nfs-layouttype4-1");
1145 
1146 module_init(nfs4filelayout_init);
1147 module_exit(nfs4filelayout_exit);
1148