pnfs.c

   1/*
   2 *  pNFS functions to call and manage layout drivers.
   3 *
   4 *  Copyright (c) 2002 [year of first publication]
   5 *  The Regents of the University of Michigan
   6 *  All Rights Reserved
   7 *
   8 *  Dean Hildebrand <dhildebz@umich.edu>
   9 *
  10 *  Permission is granted to use, copy, create derivative works, and
  11 *  redistribute this software and such derivative works for any purpose,
  12 *  so long as the name of the University of Michigan is not used in
  13 *  any advertising or publicity pertaining to the use or distribution
  14 *  of this software without specific, written prior authorization. If
  15 *  the above copyright notice or any other identification of the
  16 *  University of Michigan is included in any copy of any portion of
  17 *  this software, then the disclaimer below must also be included.
  18 *
  19 *  This software is provided as is, without representation or warranty
  20 *  of any kind either express or implied, including without limitation
  21 *  the implied warranties of merchantability, fitness for a particular
  22 *  purpose, or noninfringement.  The Regents of the University of
  23 *  Michigan shall not be liable for any damages, including special,
  24 *  indirect, incidental, or consequential damages, with respect to any
  25 *  claim arising out of or in connection with the use of the software,
  26 *  even if it has been or is hereafter advised of the possibility of
  27 *  such damages.
  28 */
  29
  30#include <linux/nfs_fs.h>
  31#include <linux/nfs_page.h>
  32#include <linux/module.h>
  33#include <linux/sort.h>
  34#include "internal.h"
  35#include "pnfs.h"
  36#include "iostat.h"
  37#include "nfs4trace.h"
  38#include "delegation.h"
  39#include "nfs42.h"
  40#include "nfs4_fs.h"
  41
  42#define NFSDBG_FACILITY		NFSDBG_PNFS
  43#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  44
  45/* Locking:
  46 *
  47 * pnfs_spinlock:
  48 *      protects pnfs_modules_tbl.
  49 */
  50static DEFINE_SPINLOCK(pnfs_spinlock);
  51
  52/*
  53 * pnfs_modules_tbl holds all pnfs modules
  54 */
  55static LIST_HEAD(pnfs_modules_tbl);
  56
  57static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  58static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
  59		struct list_head *free_me,
  60		const struct pnfs_layout_range *range,
  61		u32 seq);
  62static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
  63		                struct list_head *tmp_list);
  64static int pnfs_layout_return_on_reboot(struct pnfs_layout_hdr *lo);
  65
  66/* Return the registered pnfs layout driver module matching given id */
  67static struct pnfs_layoutdriver_type *
  68find_pnfs_driver_locked(u32 id)
  69{
  70	struct pnfs_layoutdriver_type *local;
  71
  72	list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  73		if (local->id == id)
  74			goto out;
  75	local = NULL;
  76out:
  77	dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  78	return local;
  79}
  80
  81static struct pnfs_layoutdriver_type *
  82find_pnfs_driver(u32 id)
  83{
  84	struct pnfs_layoutdriver_type *local;
  85
  86	spin_lock(&pnfs_spinlock);
  87	local = find_pnfs_driver_locked(id);
  88	if (local != NULL && !try_module_get(local->owner)) {
  89		dprintk("%s: Could not grab reference on module\n", __func__);
  90		local = NULL;
  91	}
  92	spin_unlock(&pnfs_spinlock);
  93	return local;
  94}
  95
  96const struct pnfs_layoutdriver_type *pnfs_find_layoutdriver(u32 id)
  97{
  98	return find_pnfs_driver(id);
  99}
 100
 101void pnfs_put_layoutdriver(const struct pnfs_layoutdriver_type *ld)
 102{
 103	if (ld)
 104		module_put(ld->owner);
 105}
 106
 107void
 108unset_pnfs_layoutdriver(struct nfs_server *nfss)
 109{
 110	if (nfss->pnfs_curr_ld) {
 111		if (nfss->pnfs_curr_ld->clear_layoutdriver)
 112			nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 113		/* Decrement the MDS count. Purge the deviceid cache if zero */
 114		if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 115			nfs4_deviceid_purge_client(nfss->nfs_client);
 116		module_put(nfss->pnfs_curr_ld->owner);
 117	}
 118	nfss->pnfs_curr_ld = NULL;
 119}
 120
 121/*
 122 * When the server sends a list of layout types, we choose one in the order
 123 * given in the list below.
 124 *
 125 * FIXME: should this list be configurable in some fashion? module param?
 126 * 	  mount option? something else?
 127 */
 128static const u32 ld_prefs[] = {
 129	LAYOUT_SCSI,
 130	LAYOUT_BLOCK_VOLUME,
 131	LAYOUT_OSD2_OBJECTS,
 132	LAYOUT_FLEX_FILES,
 133	LAYOUT_NFSV4_1_FILES,
 134	0
 135};
 136
 137static int
 138ld_cmp(const void *e1, const void *e2)
 139{
 140	u32 ld1 = *((u32 *)e1);
 141	u32 ld2 = *((u32 *)e2);
 142	int i;
 143
 144	for (i = 0; ld_prefs[i] != 0; i++) {
 145		if (ld1 == ld_prefs[i])
 146			return -1;
 147
 148		if (ld2 == ld_prefs[i])
 149			return 1;
 150	}
 151	return 0;
 152}
 153
 154/*
 155 * Try to set the server's pnfs module to the pnfs layout type specified by id.
 156 * Currently only one pNFS layout driver per filesystem is supported.
 157 *
 158 * @ids array of layout types supported by MDS.
 159 */
 160void
 161set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 162		      struct nfs_fsinfo *fsinfo)
 163{
 164	struct pnfs_layoutdriver_type *ld_type = NULL;
 165	u32 id;
 166	int i;
 167
 168	if (fsinfo->nlayouttypes == 0)
 169		goto out_no_driver;
 170	if (!(server->nfs_client->cl_exchange_flags &
 171		 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 172		printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 173			__func__, server->nfs_client->cl_exchange_flags);
 174		goto out_no_driver;
 175	}
 176
 177	sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 178		sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 179
 180	for (i = 0; i < fsinfo->nlayouttypes; i++) {
 181		id = fsinfo->layouttype[i];
 182		ld_type = find_pnfs_driver(id);
 183		if (!ld_type) {
 184			request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 185					id);
 186			ld_type = find_pnfs_driver(id);
 187		}
 188		if (ld_type)
 189			break;
 190	}
 191
 192	if (!ld_type) {
 193		dprintk("%s: No pNFS module found!\n", __func__);
 194		goto out_no_driver;
 195	}
 196
 197	server->pnfs_curr_ld = ld_type;
 198	if (ld_type->set_layoutdriver
 199	    && ld_type->set_layoutdriver(server, mntfh)) {
 200		printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 201			"driver %u.\n", __func__, id);
 202		module_put(ld_type->owner);
 203		goto out_no_driver;
 204	}
 205	/* Bump the MDS count */
 206	atomic_inc(&server->nfs_client->cl_mds_count);
 207
 208	dprintk("%s: pNFS module for %u set\n", __func__, id);
 209	return;
 210
 211out_no_driver:
 212	dprintk("%s: Using NFSv4 I/O\n", __func__);
 213	server->pnfs_curr_ld = NULL;
 214}
 215
 216int
 217pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 218{
 219	int status = -EINVAL;
 220	struct pnfs_layoutdriver_type *tmp;
 221
 222	if (ld_type->id == 0) {
 223		printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 224		return status;
 225	}
 226	if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 227		printk(KERN_ERR "NFS: %s Layout driver must provide "
 228		       "alloc_lseg and free_lseg.\n", __func__);
 229		return status;
 230	}
 231
 232	spin_lock(&pnfs_spinlock);
 233	tmp = find_pnfs_driver_locked(ld_type->id);
 234	if (!tmp) {
 235		list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 236		status = 0;
 237		dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 238			ld_type->name);
 239	} else {
 240		printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 241			__func__, ld_type->id);
 242	}
 243	spin_unlock(&pnfs_spinlock);
 244
 245	return status;
 246}
 247EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 248
 249void
 250pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 251{
 252	dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 253	spin_lock(&pnfs_spinlock);
 254	list_del(&ld_type->pnfs_tblid);
 255	spin_unlock(&pnfs_spinlock);
 256}
 257EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 258
 259/*
 260 * pNFS client layout cache
 261 */
 262
 263/* Need to hold i_lock if caller does not already hold reference */
 264void
 265pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 266{
 267	refcount_inc(&lo->plh_refcount);
 268}
 269
 270static struct pnfs_layout_hdr *
 271pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 272{
 273	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 274	return ld->alloc_layout_hdr(ino, gfp_flags);
 275}
 276
 277static void
 278pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 279{
 280	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 281	struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 282
 283	if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
 284		struct nfs_client *clp = server->nfs_client;
 285
 286		spin_lock(&clp->cl_lock);
 287		list_del_rcu(&lo->plh_layouts);
 288		spin_unlock(&clp->cl_lock);
 289	}
 290	put_cred(lo->plh_lc_cred);
 291	return ld->free_layout_hdr(lo);
 292}
 293
 294static void
 295pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 296{
 297	struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 298	dprintk("%s: freeing layout cache %p\n", __func__, lo);
 299	nfsi->layout = NULL;
 300	/* Reset MDS Threshold I/O counters */
 301	nfsi->write_io = 0;
 302	nfsi->read_io = 0;
 303}
 304
 305void
 306pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 307{
 308	struct inode *inode;
 309
 310	if (!lo)
 311		return;
 312	inode = lo->plh_inode;
 313	pnfs_layoutreturn_before_put_layout_hdr(lo);
 314
 315	if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 316		if (!list_empty(&lo->plh_segs))
 317			WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 318		pnfs_detach_layout_hdr(lo);
 319		/* Notify pnfs_destroy_layout_final() that we're done */
 320		if (inode_state_read(inode) & (I_FREEING | I_CLEAR))
 321			wake_up_var_locked(lo, &inode->i_lock);
 322		spin_unlock(&inode->i_lock);
 323		pnfs_free_layout_hdr(lo);
 324	}
 325}
 326
 327static struct inode *
 328pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
 329{
 330	struct inode *inode = igrab(lo->plh_inode);
 331	if (inode)
 332		return inode;
 333	set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
 334	return NULL;
 335}
 336
 337/*
 338 * Compare 2 layout stateid sequence ids, to see which is newer,
 339 * taking into account wraparound issues.
 340 */
 341static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 342{
 343	return (s32)(s1 - s2) > 0;
 344}
 345
 346static void pnfs_barrier_update(struct pnfs_layout_hdr *lo, u32 newseq)
 347{
 348	if (pnfs_seqid_is_newer(newseq, lo->plh_barrier) || !lo->plh_barrier)
 349		lo->plh_barrier = newseq;
 350}
 351
 352static void
 353pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 354			 u32 seq)
 355{
 356	if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 357		iomode = IOMODE_ANY;
 358	lo->plh_return_iomode = iomode;
 359	set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 360	/*
 361	 * We must set lo->plh_return_seq to avoid livelocks with
 362	 * pnfs_layout_need_return()
 363	 */
 364	if (seq == 0)
 365		seq = be32_to_cpu(lo->plh_stateid.seqid);
 366	if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
 367		lo->plh_return_seq = seq;
 368	pnfs_barrier_update(lo, seq);
 369}
 370
 371static void
 372pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 373{
 374	struct pnfs_layout_segment *lseg;
 375	lo->plh_return_iomode = 0;
 376	lo->plh_return_seq = 0;
 377	clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 378	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 379		if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 380			continue;
 381		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 382	}
 383}
 384
 385static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 386{
 387	clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 388	clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 389	smp_mb__after_atomic();
 390	wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 391	rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 392}
 393
 394static void
 395pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 396		struct list_head *free_me)
 397{
 398	clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 399	clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 400	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 401		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 402	if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 403		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 404}
 405
 406/*
 407 * Update the seqid of a layout stateid after receiving
 408 * NFS4ERR_OLD_STATEID
 409 */
 410bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
 411		struct pnfs_layout_range *dst_range,
 412		struct inode *inode)
 413{
 414	struct pnfs_layout_hdr *lo;
 415	struct pnfs_layout_range range = {
 416		.iomode = IOMODE_ANY,
 417		.offset = 0,
 418		.length = NFS4_MAX_UINT64,
 419	};
 420	bool ret = false;
 421	LIST_HEAD(head);
 422	int err;
 423
 424	spin_lock(&inode->i_lock);
 425	lo = NFS_I(inode)->layout;
 426	if (lo &&  pnfs_layout_is_valid(lo) &&
 427	    nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 428		/* Is our call using the most recent seqid? If so, bump it */
 429		if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
 430			nfs4_stateid_seqid_inc(dst);
 431			ret = true;
 432			goto out;
 433		}
 434		/* Try to update the seqid to the most recent */
 435		err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 436		if (err != -EBUSY) {
 437			dst->seqid = lo->plh_stateid.seqid;
 438			*dst_range = range;
 439			ret = true;
 440		}
 441	}
 442out:
 443	spin_unlock(&inode->i_lock);
 444	pnfs_free_lseg_list(&head);
 445	return ret;
 446}
 447
 448/*
 449 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 450 *
 451 * In order to continue using the pnfs_layout_hdr, a full recovery
 452 * is required.
 453 * Note that caller must hold inode->i_lock.
 454 */
 455int
 456pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 457		struct list_head *lseg_list)
 458{
 459	struct pnfs_layout_range range = {
 460		.iomode = IOMODE_ANY,
 461		.offset = 0,
 462		.length = NFS4_MAX_UINT64,
 463	};
 464	struct pnfs_layout_segment *lseg, *next;
 465
 466	if (test_and_set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
 467		return !list_empty(&lo->plh_segs);
 468	clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(lo->plh_inode)->flags);
 469	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 470		pnfs_clear_lseg_state(lseg, lseg_list);
 471	pnfs_clear_layoutreturn_info(lo);
 472	pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 473	set_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags);
 474	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 475	    !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 476		pnfs_clear_layoutreturn_waitbit(lo);
 477	return !list_empty(&lo->plh_segs);
 478}
 479
 480static int pnfs_mark_layout_stateid_return(struct pnfs_layout_hdr *lo,
 481					   struct list_head *lseg_list,
 482					   enum pnfs_iomode iomode, u32 seq)
 483{
 484	struct pnfs_layout_range range = {
 485		.iomode = iomode,
 486		.length = NFS4_MAX_UINT64,
 487	};
 488
 489	return pnfs_mark_matching_lsegs_return(lo, lseg_list, &range, seq);
 490}
 491
 492static int
 493pnfs_iomode_to_fail_bit(u32 iomode)
 494{
 495	return iomode == IOMODE_RW ?
 496		NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 497}
 498
 499static void
 500pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 501{
 502	lo->plh_retry_timestamp = jiffies;
 503	if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 504		refcount_inc(&lo->plh_refcount);
 505}
 506
 507static void
 508pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 509{
 510	if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 511		refcount_dec(&lo->plh_refcount);
 512}
 513
 514static void
 515pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 516{
 517	struct inode *inode = lo->plh_inode;
 518	struct pnfs_layout_range range = {
 519		.iomode = iomode,
 520		.offset = 0,
 521		.length = NFS4_MAX_UINT64,
 522	};
 523	LIST_HEAD(head);
 524
 525	spin_lock(&inode->i_lock);
 526	pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 527	pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 528	spin_unlock(&inode->i_lock);
 529	pnfs_free_lseg_list(&head);
 530	dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 531			iomode == IOMODE_RW ?  "RW" : "READ");
 532}
 533
 534static bool
 535pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 536{
 537	unsigned long start, end;
 538	int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 539
 540	if (test_bit(fail_bit, &lo->plh_flags) == 0)
 541		return false;
 542	end = jiffies;
 543	start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 544	if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 545		/* It is time to retry the failed layoutgets */
 546		pnfs_layout_clear_fail_bit(lo, fail_bit);
 547		return false;
 548	}
 549	return true;
 550}
 551
 552static void
 553pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 554		const struct pnfs_layout_range *range,
 555		const nfs4_stateid *stateid)
 556{
 557	INIT_LIST_HEAD(&lseg->pls_list);
 558	INIT_LIST_HEAD(&lseg->pls_lc_list);
 559	INIT_LIST_HEAD(&lseg->pls_commits);
 560	refcount_set(&lseg->pls_refcount, 1);
 561	set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 562	lseg->pls_layout = lo;
 563	lseg->pls_range = *range;
 564	lseg->pls_seq = be32_to_cpu(stateid->seqid);
 565}
 566
 567static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 568{
 569	if (lseg != NULL) {
 570		struct inode *inode = lseg->pls_layout->plh_inode;
 571		NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 572	}
 573}
 574
 575static void
 576pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 577		struct pnfs_layout_segment *lseg)
 578{
 579	WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 580	list_del_init(&lseg->pls_list);
 581	/* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 582	refcount_dec(&lo->plh_refcount);
 583	if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 584		return;
 585	if (list_empty(&lo->plh_segs) &&
 586	    !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 587	    !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 588		if (atomic_read(&lo->plh_outstanding) == 0)
 589			set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 590		clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 591	}
 592}
 593
 594static bool
 595pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 596		struct pnfs_layout_segment *lseg)
 597{
 598	if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 599	    pnfs_layout_is_valid(lo)) {
 600		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 601		list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 602		return true;
 603	}
 604	return false;
 605}
 606
 607void
 608pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 609{
 610	struct pnfs_layout_hdr *lo;
 611	struct inode *inode;
 612
 613	if (!lseg)
 614		return;
 615
 616	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 617		refcount_read(&lseg->pls_refcount),
 618		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 619
 620	lo = lseg->pls_layout;
 621	inode = lo->plh_inode;
 622
 623	if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 624		pnfs_get_layout_hdr(lo);
 625		pnfs_layout_remove_lseg(lo, lseg);
 626		if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 627			lseg = NULL;
 628		spin_unlock(&inode->i_lock);
 629		pnfs_free_lseg(lseg);
 630		pnfs_put_layout_hdr(lo);
 631	}
 632}
 633EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 634
 635/*
 636 * is l2 fully contained in l1?
 637 *   start1                             end1
 638 *   [----------------------------------)
 639 *           start2           end2
 640 *           [----------------)
 641 */
 642static bool
 643pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 644		 const struct pnfs_layout_range *l2)
 645{
 646	u64 start1 = l1->offset;
 647	u64 end1 = pnfs_end_offset(start1, l1->length);
 648	u64 start2 = l2->offset;
 649	u64 end2 = pnfs_end_offset(start2, l2->length);
 650
 651	return (start1 <= start2) && (end1 >= end2);
 652}
 653
 654static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 655		struct list_head *tmp_list)
 656{
 657	if (!refcount_dec_and_test(&lseg->pls_refcount))
 658		return false;
 659	pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 660	list_add(&lseg->pls_list, tmp_list);
 661	return true;
 662}
 663
 664/* Returns 1 if lseg is removed from list, 0 otherwise */
 665static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 666			     struct list_head *tmp_list)
 667{
 668	int rv = 0;
 669
 670	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 671		/* Remove the reference keeping the lseg in the
 672		 * list.  It will now be removed when all
 673		 * outstanding io is finished.
 674		 */
 675		dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 676			refcount_read(&lseg->pls_refcount));
 677		if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 678			rv = 1;
 679	}
 680	return rv;
 681}
 682
 683static bool
 684pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 685		 const struct pnfs_layout_range *recall_range)
 686{
 687	return (recall_range->iomode == IOMODE_ANY ||
 688		lseg_range->iomode == recall_range->iomode) &&
 689	       pnfs_lseg_range_intersecting(lseg_range, recall_range);
 690}
 691
 692static bool
 693pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 694		const struct pnfs_layout_range *recall_range,
 695		u32 seq)
 696{
 697	if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 698		return false;
 699	if (recall_range == NULL)
 700		return true;
 701	return pnfs_should_free_range(&lseg->pls_range, recall_range);
 702}
 703
 704/**
 705 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 706 * @lo: layout header containing the lsegs
 707 * @tmp_list: list head where doomed lsegs should go
 708 * @recall_range: optional recall range argument to match (may be NULL)
 709 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 710 *
 711 * Walk the list of lsegs in the layout header, and tear down any that should
 712 * be destroyed. If "recall_range" is specified then the segment must match
 713 * that range. If "seq" is non-zero, then only match segments that were handed
 714 * out at or before that sequence.
 715 *
 716 * Returns number of matching invalid lsegs remaining in list after scanning
 717 * it and purging them.
 718 */
 719int
 720pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 721			    struct list_head *tmp_list,
 722			    const struct pnfs_layout_range *recall_range,
 723			    u32 seq)
 724{
 725	struct pnfs_layout_segment *lseg, *next;
 726	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 727	int remaining = 0;
 728
 729	dprintk("%s:Begin lo %p\n", __func__, lo);
 730
 731	if (list_empty(&lo->plh_segs))
 732		return 0;
 733	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 734		if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 735			dprintk("%s: freeing lseg %p iomode %d seq %u "
 736				"offset %llu length %llu\n", __func__,
 737				lseg, lseg->pls_range.iomode, lseg->pls_seq,
 738				lseg->pls_range.offset, lseg->pls_range.length);
 739			if (mark_lseg_invalid(lseg, tmp_list))
 740				continue;
 741			remaining++;
 742			pnfs_lseg_cancel_io(server, lseg);
 743		}
 744	dprintk("%s:Return %i\n", __func__, remaining);
 745	return remaining;
 746}
 747
 748static void pnfs_reset_return_info(struct pnfs_layout_hdr *lo)
 749{
 750	struct pnfs_layout_segment *lseg;
 751
 752	list_for_each_entry(lseg, &lo->plh_return_segs, pls_list)
 753		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 754}
 755
 756static void
 757pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 758		struct list_head *free_me,
 759		const struct pnfs_layout_range *range,
 760		u32 seq)
 761{
 762	struct pnfs_layout_segment *lseg, *next;
 763
 764	list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 765		if (pnfs_match_lseg_recall(lseg, range, seq))
 766			list_move_tail(&lseg->pls_list, free_me);
 767	}
 768}
 769
 770/* note free_me must contain lsegs from a single layout_hdr */
 771void
 772pnfs_free_lseg_list(struct list_head *free_me)
 773{
 774	struct pnfs_layout_segment *lseg, *tmp;
 775
 776	if (list_empty(free_me))
 777		return;
 778
 779	list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 780		list_del(&lseg->pls_list);
 781		pnfs_free_lseg(lseg);
 782	}
 783}
 784
 785static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
 786{
 787	struct pnfs_layout_hdr *lo;
 788	LIST_HEAD(tmp_list);
 789
 790	spin_lock(&nfsi->vfs_inode.i_lock);
 791	lo = nfsi->layout;
 792	if (lo) {
 793		pnfs_get_layout_hdr(lo);
 794		pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 795		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 796		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 797		spin_unlock(&nfsi->vfs_inode.i_lock);
 798		pnfs_free_lseg_list(&tmp_list);
 799		nfs_commit_inode(&nfsi->vfs_inode, 0);
 800		pnfs_put_layout_hdr(lo);
 801	} else
 802		spin_unlock(&nfsi->vfs_inode.i_lock);
 803	return lo;
 804}
 805
 806void pnfs_destroy_layout(struct nfs_inode *nfsi)
 807{
 808	__pnfs_destroy_layout(nfsi);
 809}
 810EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 811
 812void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
 813{
 814	struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
 815	struct inode *inode = &nfsi->vfs_inode;
 816
 817	if (lo) {
 818		spin_lock(&inode->i_lock);
 819		wait_var_event_spinlock(lo, nfsi->layout != lo,
 820					&inode->i_lock);
 821		spin_unlock(&inode->i_lock);
 822	}
 823}
 824
 825static bool
 826pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 827		struct list_head *layout_list)
 828{
 829	struct pnfs_layout_hdr *lo;
 830	bool ret = false;
 831
 832	spin_lock(&inode->i_lock);
 833	lo = NFS_I(inode)->layout;
 834	if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 835		pnfs_get_layout_hdr(lo);
 836		list_add(&lo->plh_bulk_destroy, layout_list);
 837		ret = true;
 838	}
 839	spin_unlock(&inode->i_lock);
 840	return ret;
 841}
 842
 843/* Caller must hold rcu_read_lock and clp->cl_lock */
 844static int
 845pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 846		struct nfs_server *server,
 847		struct list_head *layout_list)
 848	__must_hold(&clp->cl_lock)
 849	__must_hold(RCU)
 850{
 851	struct pnfs_layout_hdr *lo, *next;
 852	struct inode *inode;
 853
 854	list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 855		if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
 856		    test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
 857		    !list_empty(&lo->plh_bulk_destroy))
 858			continue;
 859		/* If the sb is being destroyed, just bail */
 860		if (!nfs_sb_active(server->super))
 861			break;
 862		inode = pnfs_grab_inode_layout_hdr(lo);
 863		if (inode != NULL) {
 864			if (pnfs_layout_add_bulk_destroy_list(inode,
 865						layout_list))
 866				continue;
 867			rcu_read_unlock();
 868			spin_unlock(&clp->cl_lock);
 869			iput(inode);
 870		} else {
 871			rcu_read_unlock();
 872			spin_unlock(&clp->cl_lock);
 873		}
 874		nfs_sb_deactive(server->super);
 875		spin_lock(&clp->cl_lock);
 876		rcu_read_lock();
 877		return -EAGAIN;
 878	}
 879	return 0;
 880}
 881
 882static int
 883pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 884				   enum pnfs_layout_destroy_mode mode)
 885{
 886	struct pnfs_layout_hdr *lo;
 887	struct inode *inode;
 888	LIST_HEAD(lseg_list);
 889	int ret = 0;
 890
 891	while (!list_empty(layout_list)) {
 892		lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 893				plh_bulk_destroy);
 894		dprintk("%s freeing layout for inode %lu\n", __func__,
 895			lo->plh_inode->i_ino);
 896		inode = lo->plh_inode;
 897
 898		pnfs_layoutcommit_inode(inode, false);
 899
 900		spin_lock(&inode->i_lock);
 901		list_del_init(&lo->plh_bulk_destroy);
 902		if (mode == PNFS_LAYOUT_FILE_BULK_RETURN) {
 903			pnfs_mark_layout_stateid_return(lo, &lseg_list,
 904							IOMODE_ANY, 0);
 905		} else if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 906			if (mode == PNFS_LAYOUT_BULK_RETURN)
 907				set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 908			ret = -EAGAIN;
 909		}
 910		spin_unlock(&inode->i_lock);
 911		pnfs_free_lseg_list(&lseg_list);
 912		/* Free all lsegs that are attached to commit buckets */
 913		nfs_commit_inode(inode, 0);
 914		pnfs_put_layout_hdr(lo);
 915		nfs_iput_and_deactive(inode);
 916	}
 917	return ret;
 918}
 919
 920int pnfs_layout_destroy_byfsid(struct nfs_client *clp, struct nfs_fsid *fsid,
 921			       enum pnfs_layout_destroy_mode mode)
 922{
 923	struct nfs_server *server;
 924	LIST_HEAD(layout_list);
 925
 926	spin_lock(&clp->cl_lock);
 927	rcu_read_lock();
 928restart:
 929	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 930		if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 931			continue;
 932		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 933				server,
 934				&layout_list) != 0)
 935			goto restart;
 936	}
 937	rcu_read_unlock();
 938	spin_unlock(&clp->cl_lock);
 939
 940	return pnfs_layout_free_bulk_destroy_list(&layout_list, mode);
 941}
 942
 943static void pnfs_layout_build_destroy_list_byclient(struct nfs_client *clp,
 944						    struct list_head *list)
 945{
 946	struct nfs_server *server;
 947
 948	spin_lock(&clp->cl_lock);
 949	rcu_read_lock();
 950restart:
 951	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 952		if (pnfs_layout_bulk_destroy_byserver_locked(clp, server,
 953							     list) != 0)
 954			goto restart;
 955	}
 956	rcu_read_unlock();
 957	spin_unlock(&clp->cl_lock);
 958}
 959
 960static int pnfs_layout_do_destroy_byclid(struct nfs_client *clp,
 961					 struct list_head *list,
 962					 enum pnfs_layout_destroy_mode mode)
 963{
 964	pnfs_layout_build_destroy_list_byclient(clp, list);
 965	return pnfs_layout_free_bulk_destroy_list(list, mode);
 966}
 967
 968int pnfs_layout_destroy_byclid(struct nfs_client *clp,
 969			       enum pnfs_layout_destroy_mode mode)
 970{
 971	LIST_HEAD(layout_list);
 972
 973	return pnfs_layout_do_destroy_byclid(clp, &layout_list, mode);
 974}
 975
 976/*
 977 * Called by the state manager to remove all layouts established under an
 978 * expired lease.
 979 */
 980void
 981pnfs_destroy_all_layouts(struct nfs_client *clp)
 982{
 983	nfs4_deviceid_mark_client_invalid(clp);
 984	nfs4_deviceid_purge_client(clp);
 985
 986	pnfs_layout_destroy_byclid(clp, PNFS_LAYOUT_INVALIDATE);
 987}
 988
 989static void pnfs_layout_build_recover_list_byclient(struct nfs_client *clp,
 990						    struct list_head *list)
 991{
 992	struct nfs_server *server;
 993
 994	spin_lock(&clp->cl_lock);
 995	rcu_read_lock();
 996restart:
 997	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 998		if (!(server->caps & NFS_CAP_REBOOT_LAYOUTRETURN))
 999			continue;
1000		if (pnfs_layout_bulk_destroy_byserver_locked(clp, server,
1001							     list) != 0)
1002			goto restart;
1003	}
1004	rcu_read_unlock();
1005	spin_unlock(&clp->cl_lock);
1006}
1007
1008static int pnfs_layout_bulk_list_reboot(struct list_head *list)
1009{
1010	struct pnfs_layout_hdr *lo;
1011	struct nfs_server *server;
1012	int ret;
1013
1014	list_for_each_entry(lo, list, plh_bulk_destroy) {
1015		server = NFS_SERVER(lo->plh_inode);
1016		ret = pnfs_layout_return_on_reboot(lo);
1017		switch (ret) {
1018		case 0:
1019			continue;
1020		case -NFS4ERR_BAD_STATEID:
1021			server->caps &= ~NFS_CAP_REBOOT_LAYOUTRETURN;
1022			break;
1023		case -NFS4ERR_NO_GRACE:
1024			break;
1025		default:
1026			goto err;
1027		}
1028		break;
1029	}
1030	return 0;
1031err:
1032	return ret;
1033}
1034
1035int pnfs_layout_handle_reboot(struct nfs_client *clp)
1036{
1037	LIST_HEAD(list);
1038	int ret = 0, ret2;
1039
1040	pnfs_layout_build_recover_list_byclient(clp, &list);
1041	if (!list_empty(&list))
1042		ret = pnfs_layout_bulk_list_reboot(&list);
1043	ret2 = pnfs_layout_do_destroy_byclid(clp, &list,
1044					     PNFS_LAYOUT_INVALIDATE);
1045	if (!ret)
1046		ret = ret2;
1047	return (ret == 0) ?  0 : -EAGAIN;
1048}
1049
1050static void
1051pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
1052{
1053	const struct cred *old;
1054
1055	if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
1056		old = xchg(&lo->plh_lc_cred, get_cred(cred));
1057		put_cred(old);
1058	}
1059}
1060
1061/* update lo->plh_stateid with new if is more recent */
1062void
1063pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
1064			const struct cred *cred, bool update_barrier)
1065{
1066	u32 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
1067	u32 newseq = be32_to_cpu(new->seqid);
1068
1069	if (!pnfs_layout_is_valid(lo)) {
1070		pnfs_set_layout_cred(lo, cred);
1071		nfs4_stateid_copy(&lo->plh_stateid, new);
1072		lo->plh_barrier = newseq;
1073		pnfs_clear_layoutreturn_info(lo);
1074		clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1075		return;
1076	}
1077
1078	if (pnfs_seqid_is_newer(newseq, oldseq))
1079		nfs4_stateid_copy(&lo->plh_stateid, new);
1080
1081	if (update_barrier) {
1082		pnfs_barrier_update(lo, newseq);
1083		return;
1084	}
1085	/*
1086	 * Because of wraparound, we want to keep the barrier
1087	 * "close" to the current seqids. We really only want to
1088	 * get here from a layoutget call.
1089	 */
1090	if (atomic_read(&lo->plh_outstanding) == 1)
1091		 pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
1092}
1093
1094static bool
1095pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
1096		const nfs4_stateid *stateid)
1097{
1098	u32 seqid = be32_to_cpu(stateid->seqid);
1099
1100	return lo->plh_barrier && pnfs_seqid_is_newer(lo->plh_barrier, seqid);
1101}
1102
1103/* lget is set to 1 if called from inside send_layoutget call chain */
1104static bool
1105pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1106{
1107	return lo->plh_block_lgets ||
1108		test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1109}
1110
1111static struct nfs_server *
1112pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1113{
1114	struct nfs_server *server;
1115
1116	if (inode) {
1117		server = NFS_SERVER(inode);
1118	} else {
1119		struct dentry *parent_dir = dget_parent(ctx->dentry);
1120		server = NFS_SERVER(parent_dir->d_inode);
1121		dput(parent_dir);
1122	}
1123	return server;
1124}
1125
1126static void nfs4_free_pages(struct page **pages, size_t size)
1127{
1128	int i;
1129
1130	if (!pages)
1131		return;
1132
1133	for (i = 0; i < size; i++) {
1134		if (!pages[i])
1135			break;
1136		__free_page(pages[i]);
1137	}
1138	kfree(pages);
1139}
1140
1141static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1142{
1143	struct page **pages;
1144	int i;
1145
1146	pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1147	if (!pages) {
1148		dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1149		return NULL;
1150	}
1151
1152	for (i = 0; i < size; i++) {
1153		pages[i] = alloc_page(gfp_flags);
1154		if (!pages[i]) {
1155			dprintk("%s: failed to allocate page\n", __func__);
1156			nfs4_free_pages(pages, i);
1157			return NULL;
1158		}
1159	}
1160
1161	return pages;
1162}
1163
1164static struct nfs4_layoutget *
1165pnfs_alloc_init_layoutget_args(struct inode *ino,
1166	   struct nfs_open_context *ctx,
1167	   const nfs4_stateid *stateid,
1168	   const struct pnfs_layout_range *range,
1169	   gfp_t gfp_flags)
1170{
1171	struct nfs_server *server = pnfs_find_server(ino, ctx);
1172	size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1173	size_t max_pages = max_response_pages(server);
1174	struct nfs4_layoutget *lgp;
1175
1176	dprintk("--> %s\n", __func__);
1177
1178	lgp = kzalloc(sizeof(*lgp), gfp_flags);
1179	if (lgp == NULL)
1180		return NULL;
1181
1182	if (max_reply_sz) {
1183		size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1184		if (npages < max_pages)
1185			max_pages = npages;
1186	}
1187
1188	lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1189	if (!lgp->args.layout.pages) {
1190		kfree(lgp);
1191		return NULL;
1192	}
1193	lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1194	lgp->res.layoutp = &lgp->args.layout;
1195
1196	/* Don't confuse uninitialised result and success */
1197	lgp->res.status = -NFS4ERR_DELAY;
1198
1199	lgp->args.minlength = PAGE_SIZE;
1200	if (lgp->args.minlength > range->length)
1201		lgp->args.minlength = range->length;
1202	if (ino) {
1203		loff_t i_size = i_size_read(ino);
1204
1205		if (range->iomode == IOMODE_READ) {
1206			if (range->offset >= i_size)
1207				lgp->args.minlength = 0;
1208			else if (i_size - range->offset < lgp->args.minlength)
1209				lgp->args.minlength = i_size - range->offset;
1210		}
1211	}
1212	lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1213	pnfs_copy_range(&lgp->args.range, range);
1214	lgp->args.type = server->pnfs_curr_ld->id;
1215	lgp->args.inode = ino;
1216	lgp->args.ctx = get_nfs_open_context(ctx);
1217	nfs4_stateid_copy(&lgp->args.stateid, stateid);
1218	lgp->gfp_flags = gfp_flags;
1219	lgp->cred = ctx->cred;
1220	return lgp;
1221}
1222
1223void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1224{
1225	size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1226
1227	nfs4_free_pages(lgp->args.layout.pages, max_pages);
1228	pnfs_put_layout_hdr(lgp->lo);
1229	put_nfs_open_context(lgp->args.ctx);
1230	kfree(lgp);
1231}
1232
1233static void pnfs_clear_layoutcommit(struct inode *inode,
1234		struct list_head *head)
1235{
1236	struct nfs_inode *nfsi = NFS_I(inode);
1237	struct pnfs_layout_segment *lseg, *tmp;
1238
1239	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1240		return;
1241	list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1242		if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1243			continue;
1244		pnfs_lseg_dec_and_remove_zero(lseg, head);
1245	}
1246}
1247
1248static void
1249pnfs_layoutreturn_retry_later_locked(struct pnfs_layout_hdr *lo,
1250				     const nfs4_stateid *arg_stateid,
1251				     const struct pnfs_layout_range *range,
1252				     struct list_head *freeme)
1253{
1254	if (pnfs_layout_is_valid(lo) &&
1255	    nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1256		pnfs_reset_return_info(lo);
1257	else
1258		pnfs_mark_layout_stateid_invalid(lo, freeme);
1259	pnfs_clear_layoutreturn_waitbit(lo);
1260}
1261
1262void pnfs_layoutreturn_retry_later(struct pnfs_layout_hdr *lo,
1263				   const nfs4_stateid *arg_stateid,
1264				   const struct pnfs_layout_range *range)
1265{
1266	struct inode *inode = lo->plh_inode;
1267	LIST_HEAD(freeme);
1268
1269	spin_lock(&inode->i_lock);
1270	pnfs_layoutreturn_retry_later_locked(lo, arg_stateid, range, &freeme);
1271	spin_unlock(&inode->i_lock);
1272	pnfs_free_lseg_list(&freeme);
1273}
1274
1275void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1276		const nfs4_stateid *arg_stateid,
1277		const struct pnfs_layout_range *range,
1278		const nfs4_stateid *stateid)
1279{
1280	struct inode *inode = lo->plh_inode;
1281	LIST_HEAD(freeme);
1282
1283	spin_lock(&inode->i_lock);
1284	if (!nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1285		goto out_unlock;
1286	if (stateid && pnfs_layout_is_valid(lo)) {
1287		u32 seq = be32_to_cpu(arg_stateid->seqid);
1288
1289		pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1290		pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1291		pnfs_set_layout_stateid(lo, stateid, NULL, true);
1292		pnfs_reset_return_info(lo);
1293	} else
1294		pnfs_mark_layout_stateid_invalid(lo, &freeme);
1295out_unlock:
1296	pnfs_clear_layoutreturn_waitbit(lo);
1297	spin_unlock(&inode->i_lock);
1298	pnfs_free_lseg_list(&freeme);
1299
1300}
1301
1302static bool
1303pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1304		nfs4_stateid *stateid,
1305		const struct cred **cred,
1306		enum pnfs_iomode *iomode)
1307{
1308	/* Serialise LAYOUTGET/LAYOUTRETURN */
1309	if (atomic_read(&lo->plh_outstanding) != 0 && lo->plh_return_seq == 0)
1310		return false;
1311	if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1312		return false;
1313	set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1314	pnfs_get_layout_hdr(lo);
1315	nfs4_stateid_copy(stateid, &lo->plh_stateid);
1316	*cred = get_cred(lo->plh_lc_cred);
1317	if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1318		if (lo->plh_return_seq != 0)
1319			stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1320		if (iomode != NULL)
1321			*iomode = lo->plh_return_iomode;
1322		pnfs_clear_layoutreturn_info(lo);
1323	} else if (iomode != NULL)
1324		*iomode = IOMODE_ANY;
1325	pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1326	return true;
1327}
1328
1329static void
1330pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1331		struct pnfs_layout_hdr *lo,
1332		const nfs4_stateid *stateid,
1333		enum pnfs_iomode iomode)
1334{
1335	struct inode *inode = lo->plh_inode;
1336
1337	args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1338	args->inode = inode;
1339	args->range.iomode = iomode;
1340	args->range.offset = 0;
1341	args->range.length = NFS4_MAX_UINT64;
1342	args->layout = lo;
1343	nfs4_stateid_copy(&args->stateid, stateid);
1344}
1345
1346static int
1347pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1348		       const nfs4_stateid *stateid,
1349		       const struct cred **pcred,
1350		       enum pnfs_iomode iomode,
1351		       unsigned int flags)
1352{
1353	struct inode *ino = lo->plh_inode;
1354	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1355	struct nfs4_layoutreturn *lrp;
1356	const struct cred *cred = *pcred;
1357	int status = 0;
1358
1359	*pcred = NULL;
1360	lrp = kzalloc(sizeof(*lrp), nfs_io_gfp_mask());
1361	if (unlikely(lrp == NULL)) {
1362		status = -ENOMEM;
1363		spin_lock(&ino->i_lock);
1364		pnfs_clear_layoutreturn_waitbit(lo);
1365		spin_unlock(&ino->i_lock);
1366		put_cred(cred);
1367		pnfs_put_layout_hdr(lo);
1368		goto out;
1369	}
1370
1371	pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1372	lrp->args.ld_private = &lrp->ld_private;
1373	lrp->clp = NFS_SERVER(ino)->nfs_client;
1374	lrp->cred = cred;
1375	if (ld->prepare_layoutreturn)
1376		ld->prepare_layoutreturn(&lrp->args);
1377
1378	status = nfs4_proc_layoutreturn(lrp, flags);
1379out:
1380	dprintk("<-- %s status: %d\n", __func__, status);
1381	return status;
1382}
1383
1384/* Return true if layoutreturn is needed */
1385static bool
1386pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1387{
1388	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1389		return false;
1390	return pnfs_mark_layout_stateid_return(lo, &lo->plh_return_segs,
1391					       lo->plh_return_iomode,
1392					       lo->plh_return_seq) != EBUSY;
1393}
1394
1395static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1396{
1397	struct inode *inode= lo->plh_inode;
1398
1399	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1400		return;
1401	spin_lock(&inode->i_lock);
1402	if (pnfs_layout_need_return(lo)) {
1403		const struct cred *cred;
1404		nfs4_stateid stateid;
1405		enum pnfs_iomode iomode;
1406		bool send;
1407
1408		send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1409		spin_unlock(&inode->i_lock);
1410		if (send) {
1411			/* Send an async layoutreturn so we dont deadlock */
1412			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
1413					       PNFS_FL_LAYOUTRETURN_ASYNC);
1414		}
1415	} else
1416		spin_unlock(&inode->i_lock);
1417}
1418
1419/*
1420 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1421 * when the layout segment list is empty.
1422 *
1423 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1424 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1425 * deviceid is marked invalid.
1426 */
1427int
1428_pnfs_return_layout(struct inode *ino)
1429{
1430	struct pnfs_layout_hdr *lo = NULL;
1431	struct nfs_inode *nfsi = NFS_I(ino);
1432	struct pnfs_layout_range range = {
1433		.iomode		= IOMODE_ANY,
1434		.offset		= 0,
1435		.length		= NFS4_MAX_UINT64,
1436	};
1437	LIST_HEAD(tmp_list);
1438	const struct cred *cred;
1439	nfs4_stateid stateid;
1440	int status = 0;
1441	bool send, valid_layout;
1442
1443	dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1444
1445	spin_lock(&ino->i_lock);
1446	lo = nfsi->layout;
1447	if (!lo) {
1448		spin_unlock(&ino->i_lock);
1449		dprintk("NFS: %s no layout to return\n", __func__);
1450		goto out;
1451	}
1452	/* Reference matched in nfs4_layoutreturn_release */
1453	pnfs_get_layout_hdr(lo);
1454	/* Is there an outstanding layoutreturn ? */
1455	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1456		spin_unlock(&ino->i_lock);
1457		if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1458					TASK_UNINTERRUPTIBLE))
1459			goto out_put_layout_hdr;
1460		spin_lock(&ino->i_lock);
1461	}
1462	valid_layout = pnfs_layout_is_valid(lo);
1463	pnfs_clear_layoutcommit(ino, &tmp_list);
1464	pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1465
1466	if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1467		NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1468
1469	/* Don't send a LAYOUTRETURN if list was initially empty */
1470	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1471			!valid_layout) {
1472		spin_unlock(&ino->i_lock);
1473		dprintk("NFS: %s no layout segments to return\n", __func__);
1474		goto out_wait_layoutreturn;
1475	}
1476
1477	send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1478	spin_unlock(&ino->i_lock);
1479	if (send)
1480		status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY,
1481						0);
1482out_wait_layoutreturn:
1483	wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1484out_put_layout_hdr:
1485	pnfs_free_lseg_list(&tmp_list);
1486	pnfs_put_layout_hdr(lo);
1487out:
1488	dprintk("<-- %s status: %d\n", __func__, status);
1489	return status;
1490}
1491
1492int
1493pnfs_commit_and_return_layout(struct inode *inode)
1494{
1495	struct pnfs_layout_hdr *lo;
1496	int ret;
1497
1498	spin_lock(&inode->i_lock);
1499	lo = NFS_I(inode)->layout;
1500	if (lo == NULL) {
1501		spin_unlock(&inode->i_lock);
1502		return 0;
1503	}
1504	pnfs_get_layout_hdr(lo);
1505	/* Block new layoutgets and read/write to ds */
1506	lo->plh_block_lgets++;
1507	spin_unlock(&inode->i_lock);
1508	filemap_fdatawait(inode->i_mapping);
1509	ret = pnfs_layoutcommit_inode(inode, true);
1510	if (ret == 0)
1511		ret = _pnfs_return_layout(inode);
1512	spin_lock(&inode->i_lock);
1513	lo->plh_block_lgets--;
1514	spin_unlock(&inode->i_lock);
1515	pnfs_put_layout_hdr(lo);
1516	return ret;
1517}
1518
1519static int pnfs_layout_return_on_reboot(struct pnfs_layout_hdr *lo)
1520{
1521	struct inode *inode = lo->plh_inode;
1522	const struct cred *cred;
1523
1524	spin_lock(&inode->i_lock);
1525	if (!pnfs_layout_is_valid(lo)) {
1526		spin_unlock(&inode->i_lock);
1527		return 0;
1528	}
1529	cred = get_cred(lo->plh_lc_cred);
1530	pnfs_get_layout_hdr(lo);
1531	spin_unlock(&inode->i_lock);
1532
1533	return pnfs_send_layoutreturn(lo, &zero_stateid, &cred, IOMODE_ANY,
1534				      PNFS_FL_LAYOUTRETURN_PRIVILEGED);
1535}
1536
1537bool pnfs_roc(struct inode *ino, struct nfs4_layoutreturn_args *args,
1538	      struct nfs4_layoutreturn_res *res, const struct cred *cred,
1539	      bool sync)
1540{
1541	struct nfs_inode *nfsi = NFS_I(ino);
1542	struct nfs_open_context *ctx;
1543	struct nfs4_state *state;
1544	struct pnfs_layout_hdr *lo;
1545	struct pnfs_layout_segment *lseg, *next;
1546	const struct cred *lc_cred;
1547	nfs4_stateid stateid;
1548	enum pnfs_iomode iomode = 0;
1549	bool layoutreturn = false, roc = false;
1550	bool skip_read;
1551
1552	if (!nfs_have_layout(ino))
1553		return false;
1554retry:
1555	rcu_read_lock();
1556	spin_lock(&ino->i_lock);
1557	lo = nfsi->layout;
1558	if (!lo || !pnfs_layout_is_valid(lo) ||
1559	    test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1560		lo = NULL;
1561		goto out_noroc;
1562	}
1563
1564	/* no roc if we hold a delegation */
1565	skip_read = false;
1566	if (nfs4_check_delegation(ino, FMODE_READ)) {
1567		if (nfs4_check_delegation(ino, FMODE_WRITE)) {
1568			lo = NULL;
1569			goto out_noroc;
1570		}
1571		skip_read = true;
1572	}
1573
1574	list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1575		state = ctx->state;
1576		if (state == NULL)
1577			continue;
1578		/* Don't return layout if there is open file state */
1579		if (state->state & FMODE_WRITE) {
1580			lo = NULL;
1581			goto out_noroc;
1582		}
1583		if (state->state & FMODE_READ)
1584			skip_read = true;
1585	}
1586
1587	if (skip_read) {
1588		bool writes = false;
1589
1590		list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1591			if (lseg->pls_range.iomode != IOMODE_READ) {
1592				writes = true;
1593				break;
1594			}
1595		}
1596		if (!writes) {
1597			lo = NULL;
1598			goto out_noroc;
1599		}
1600	}
1601
1602	pnfs_get_layout_hdr(lo);
1603	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1604		if (!sync) {
1605			pnfs_set_plh_return_info(
1606				lo, skip_read ? IOMODE_RW : IOMODE_ANY, 0);
1607			goto out_noroc;
1608		}
1609		spin_unlock(&ino->i_lock);
1610		rcu_read_unlock();
1611		wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1612			    TASK_UNINTERRUPTIBLE);
1613		pnfs_put_layout_hdr(lo);
1614		goto retry;
1615	}
1616
1617	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1618		if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1619			continue;
1620		/* If we are sending layoutreturn, invalidate all valid lsegs */
1621		if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1622			continue;
1623		/*
1624		 * Note: mark lseg for return so pnfs_layout_remove_lseg
1625		 * doesn't invalidate the layout for us.
1626		 */
1627		set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1628		if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1629			continue;
1630		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1631	}
1632
1633	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1634		goto out_noroc;
1635
1636	/* ROC in two conditions:
1637	 * 1. there are ROC lsegs
1638	 * 2. we don't send layoutreturn
1639	 */
1640	/* lo ref dropped in pnfs_roc_release() */
1641	layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1642	/* If the creds don't match, we can't compound the layoutreturn */
1643	if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1644		goto out_noroc;
1645
1646	roc = layoutreturn;
1647	pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1648	res->lrs_present = 0;
1649	layoutreturn = false;
1650	put_cred(lc_cred);
1651
1652out_noroc:
1653	spin_unlock(&ino->i_lock);
1654	rcu_read_unlock();
1655	pnfs_layoutcommit_inode(ino, sync);
1656	if (roc) {
1657		struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1658		if (ld->prepare_layoutreturn)
1659			ld->prepare_layoutreturn(args);
1660		pnfs_put_layout_hdr(lo);
1661		return true;
1662	}
1663	if (layoutreturn)
1664		pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, 0);
1665	pnfs_put_layout_hdr(lo);
1666	return false;
1667}
1668
1669int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1670		  struct nfs4_layoutreturn_res **respp, int *ret)
1671{
1672	struct nfs4_layoutreturn_args *arg = *argpp;
1673	int retval = -EAGAIN;
1674
1675	if (!arg)
1676		return 0;
1677	/* Handle Layoutreturn errors */
1678	switch (*ret) {
1679	case 0:
1680		retval = 0;
1681		break;
1682	case -NFS4ERR_NOMATCHING_LAYOUT:
1683		/* Was there an RPC level error? If not, retry */
1684		if (task->tk_rpc_status == 0)
1685			break;
1686		/*
1687		 * Is there a fatal network level error?
1688		 * If so release the layout, but flag the error.
1689		 */
1690		if ((task->tk_rpc_status == -ENETDOWN ||
1691		     task->tk_rpc_status == -ENETUNREACH) &&
1692		    task->tk_flags & RPC_TASK_NETUNREACH_FATAL) {
1693			*ret = 0;
1694			(*respp)->lrs_present = 0;
1695			retval = -EIO;
1696			break;
1697		}
1698		/* If the call was not sent, let caller handle it */
1699		if (!RPC_WAS_SENT(task))
1700			return 0;
1701		/*
1702		 * Otherwise, assume the call succeeded and
1703		 * that we need to release the layout
1704		 */
1705		*ret = 0;
1706		(*respp)->lrs_present = 0;
1707		retval = 0;
1708		break;
1709	case -NFS4ERR_DELAY:
1710		/* Let the caller handle the retry */
1711		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1712		return 0;
1713	case -NFS4ERR_OLD_STATEID:
1714		if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1715						     &arg->range, arg->inode))
1716			break;
1717		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1718		return -EAGAIN;
1719	}
1720	*argpp = NULL;
1721	*respp = NULL;
1722	return retval;
1723}
1724
1725void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1726		      struct nfs4_layoutreturn_res *res, int ret)
1727{
1728	struct pnfs_layout_hdr *lo = args->layout;
1729	struct inode *inode = args->inode;
1730	const nfs4_stateid *res_stateid = NULL;
1731	struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1732	LIST_HEAD(freeme);
1733
1734	switch (ret) {
1735	case -NFS4ERR_BADSESSION:
1736	case -NFS4ERR_DEADSESSION:
1737	case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1738	case -NFS4ERR_NOMATCHING_LAYOUT:
1739		spin_lock(&inode->i_lock);
1740		pnfs_layoutreturn_retry_later_locked(lo, &args->stateid,
1741						     &args->range, &freeme);
1742		spin_unlock(&inode->i_lock);
1743		pnfs_free_lseg_list(&freeme);
1744		break;
1745	case 0:
1746		if (res->lrs_present)
1747			res_stateid = &res->stateid;
1748		fallthrough;
1749	default:
1750		pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1751					     res_stateid);
1752	}
1753	trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1754	if (ld_private && ld_private->ops && ld_private->ops->free)
1755		ld_private->ops->free(ld_private);
1756	pnfs_put_layout_hdr(lo);
1757}
1758
1759bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1760{
1761	struct nfs_inode *nfsi = NFS_I(ino);
1762        struct pnfs_layout_hdr *lo;
1763        bool sleep = false;
1764
1765	/* we might not have grabbed lo reference. so need to check under
1766	 * i_lock */
1767        spin_lock(&ino->i_lock);
1768        lo = nfsi->layout;
1769        if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1770                rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1771                sleep = true;
1772	}
1773        spin_unlock(&ino->i_lock);
1774        return sleep;
1775}
1776
1777/*
1778 * Compare two layout segments for sorting into layout cache.
1779 * We want to preferentially return RW over RO layouts, so ensure those
1780 * are seen first.
1781 */
1782static s64
1783pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1784	   const struct pnfs_layout_range *l2)
1785{
1786	s64 d;
1787
1788	/* high offset > low offset */
1789	d = l1->offset - l2->offset;
1790	if (d)
1791		return d;
1792
1793	/* short length > long length */
1794	d = l2->length - l1->length;
1795	if (d)
1796		return d;
1797
1798	/* read > read/write */
1799	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1800}
1801
1802static bool
1803pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1804		const struct pnfs_layout_range *l2)
1805{
1806	return pnfs_lseg_range_cmp(l1, l2) > 0;
1807}
1808
1809static bool
1810pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1811		struct pnfs_layout_segment *old)
1812{
1813	return false;
1814}
1815
1816void
1817pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1818		   struct pnfs_layout_segment *lseg,
1819		   bool (*is_after)(const struct pnfs_layout_range *,
1820			   const struct pnfs_layout_range *),
1821		   bool (*do_merge)(struct pnfs_layout_segment *,
1822			   struct pnfs_layout_segment *),
1823		   struct list_head *free_me)
1824{
1825	struct pnfs_layout_segment *lp, *tmp;
1826
1827	dprintk("%s:Begin\n", __func__);
1828
1829	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1830		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1831			continue;
1832		if (do_merge(lseg, lp)) {
1833			mark_lseg_invalid(lp, free_me);
1834			continue;
1835		}
1836		if (is_after(&lseg->pls_range, &lp->pls_range))
1837			continue;
1838		list_add_tail(&lseg->pls_list, &lp->pls_list);
1839		dprintk("%s: inserted lseg %p "
1840			"iomode %d offset %llu length %llu before "
1841			"lp %p iomode %d offset %llu length %llu\n",
1842			__func__, lseg, lseg->pls_range.iomode,
1843			lseg->pls_range.offset, lseg->pls_range.length,
1844			lp, lp->pls_range.iomode, lp->pls_range.offset,
1845			lp->pls_range.length);
1846		goto out;
1847	}
1848	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1849	dprintk("%s: inserted lseg %p "
1850		"iomode %d offset %llu length %llu at tail\n",
1851		__func__, lseg, lseg->pls_range.iomode,
1852		lseg->pls_range.offset, lseg->pls_range.length);
1853out:
1854	pnfs_get_layout_hdr(lo);
1855
1856	dprintk("%s:Return\n", __func__);
1857}
1858EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1859
1860static void
1861pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1862		   struct pnfs_layout_segment *lseg,
1863		   struct list_head *free_me)
1864{
1865	struct inode *inode = lo->plh_inode;
1866	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1867
1868	if (ld->add_lseg != NULL)
1869		ld->add_lseg(lo, lseg, free_me);
1870	else
1871		pnfs_generic_layout_insert_lseg(lo, lseg,
1872				pnfs_lseg_range_is_after,
1873				pnfs_lseg_no_merge,
1874				free_me);
1875}
1876
1877static struct pnfs_layout_hdr *
1878alloc_init_layout_hdr(struct inode *ino,
1879		      struct nfs_open_context *ctx,
1880		      gfp_t gfp_flags)
1881{
1882	struct pnfs_layout_hdr *lo;
1883
1884	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1885	if (!lo)
1886		return NULL;
1887	refcount_set(&lo->plh_refcount, 1);
1888	INIT_LIST_HEAD(&lo->plh_layouts);
1889	INIT_LIST_HEAD(&lo->plh_segs);
1890	INIT_LIST_HEAD(&lo->plh_return_segs);
1891	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1892	lo->plh_inode = ino;
1893	lo->plh_lc_cred = get_cred(ctx->cred);
1894	lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1895	return lo;
1896}
1897
1898static struct pnfs_layout_hdr *
1899pnfs_find_alloc_layout(struct inode *ino,
1900		       struct nfs_open_context *ctx,
1901		       gfp_t gfp_flags)
1902	__releases(&ino->i_lock)
1903	__acquires(&ino->i_lock)
1904{
1905	struct nfs_inode *nfsi = NFS_I(ino);
1906	struct pnfs_layout_hdr *new = NULL;
1907
1908	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1909
1910	if (nfsi->layout != NULL)
1911		goto out_existing;
1912	spin_unlock(&ino->i_lock);
1913	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1914	spin_lock(&ino->i_lock);
1915
1916	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1917		nfsi->layout = new;
1918		return new;
1919	} else if (new != NULL)
1920		pnfs_free_layout_hdr(new);
1921out_existing:
1922	pnfs_get_layout_hdr(nfsi->layout);
1923	return nfsi->layout;
1924}
1925
1926/*
1927 * iomode matching rules:
1928 * iomode	lseg	strict match
1929 *                      iomode
1930 * -----	-----	------ -----
1931 * ANY		READ	N/A    true
1932 * ANY		RW	N/A    true
1933 * RW		READ	N/A    false
1934 * RW		RW	N/A    true
1935 * READ		READ	N/A    true
1936 * READ		RW	true   false
1937 * READ		RW	false  true
1938 */
1939static bool
1940pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1941		 const struct pnfs_layout_range *range,
1942		 bool strict_iomode)
1943{
1944	struct pnfs_layout_range range1;
1945
1946	if ((range->iomode == IOMODE_RW &&
1947	     ls_range->iomode != IOMODE_RW) ||
1948	    (range->iomode != ls_range->iomode &&
1949	     strict_iomode) ||
1950	    !pnfs_lseg_range_intersecting(ls_range, range))
1951		return false;
1952
1953	/* range1 covers only the first byte in the range */
1954	range1 = *range;
1955	range1.length = 1;
1956	return pnfs_lseg_range_contained(ls_range, &range1);
1957}
1958
1959/*
1960 * lookup range in layout
1961 */
1962static struct pnfs_layout_segment *
1963pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1964		struct pnfs_layout_range *range,
1965		bool strict_iomode)
1966{
1967	struct pnfs_layout_segment *lseg, *ret = NULL;
1968
1969	dprintk("%s:Begin\n", __func__);
1970
1971	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1972		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1973		    pnfs_lseg_range_match(&lseg->pls_range, range,
1974					  strict_iomode)) {
1975			ret = pnfs_get_lseg(lseg);
1976			break;
1977		}
1978	}
1979
1980	dprintk("%s:Return lseg %p ref %d\n",
1981		__func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1982	return ret;
1983}
1984
1985/*
1986 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1987 * to the MDS or over pNFS
1988 *
1989 * The nfs_inode read_io and write_io fields are cumulative counters reset
1990 * when there are no layout segments. Note that in pnfs_update_layout iomode
1991 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1992 * WRITE request.
1993 *
1994 * A return of true means use MDS I/O.
1995 *
1996 * From rfc 5661:
1997 * If a file's size is smaller than the file size threshold, data accesses
1998 * SHOULD be sent to the metadata server.  If an I/O request has a length that
1999 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
2000 * server.  If both file size and I/O size are provided, the client SHOULD
2001 * reach or exceed  both thresholds before sending its read or write
2002 * requests to the data server.
2003 */
2004static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
2005				     struct inode *ino, int iomode)
2006{
2007	struct nfs4_threshold *t = ctx->mdsthreshold;
2008	struct nfs_inode *nfsi = NFS_I(ino);
2009	loff_t fsize = i_size_read(ino);
2010	bool size = false, size_set = false, io = false, io_set = false, ret = false;
2011
2012	if (t == NULL)
2013		return ret;
2014
2015	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
2016		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
2017
2018	switch (iomode) {
2019	case IOMODE_READ:
2020		if (t->bm & THRESHOLD_RD) {
2021			dprintk("%s fsize %llu\n", __func__, fsize);
2022			size_set = true;
2023			if (fsize < t->rd_sz)
2024				size = true;
2025		}
2026		if (t->bm & THRESHOLD_RD_IO) {
2027			dprintk("%s nfsi->read_io %llu\n", __func__,
2028				nfsi->read_io);
2029			io_set = true;
2030			if (nfsi->read_io < t->rd_io_sz)
2031				io = true;
2032		}
2033		break;
2034	case IOMODE_RW:
2035		if (t->bm & THRESHOLD_WR) {
2036			dprintk("%s fsize %llu\n", __func__, fsize);
2037			size_set = true;
2038			if (fsize < t->wr_sz)
2039				size = true;
2040		}
2041		if (t->bm & THRESHOLD_WR_IO) {
2042			dprintk("%s nfsi->write_io %llu\n", __func__,
2043				nfsi->write_io);
2044			io_set = true;
2045			if (nfsi->write_io < t->wr_io_sz)
2046				io = true;
2047		}
2048		break;
2049	}
2050	if (size_set && io_set) {
2051		if (size && io)
2052			ret = true;
2053	} else if (size || io)
2054		ret = true;
2055
2056	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
2057	return ret;
2058}
2059
2060static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
2061{
2062	/*
2063	 * send layoutcommit as it can hold up layoutreturn due to lseg
2064	 * reference
2065	 */
2066	pnfs_layoutcommit_inode(lo->plh_inode, false);
2067	return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
2068				   nfs_wait_bit_killable,
2069				   TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
2070}
2071
2072static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
2073{
2074	atomic_inc(&lo->plh_outstanding);
2075}
2076
2077static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
2078{
2079	if (atomic_dec_and_test(&lo->plh_outstanding) &&
2080	    test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags)) {
2081		smp_mb__after_atomic();
2082		wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
2083	}
2084}
2085
2086static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
2087{
2088	return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
2089}
2090
2091static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
2092{
2093	unsigned long *bitlock = &lo->plh_flags;
2094
2095	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
2096	smp_mb__after_atomic();
2097	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
2098}
2099
2100static void _add_to_server_list(struct pnfs_layout_hdr *lo,
2101				struct nfs_server *server)
2102{
2103	if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
2104		struct nfs_client *clp = server->nfs_client;
2105
2106		/* The lo must be on the clp list if there is any
2107		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
2108		 */
2109		spin_lock(&clp->cl_lock);
2110		list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
2111		spin_unlock(&clp->cl_lock);
2112	}
2113}
2114
2115/*
2116 * Layout segment is retreived from the server if not cached.
2117 * The appropriate layout segment is referenced and returned to the caller.
2118 */
2119struct pnfs_layout_segment *
2120pnfs_update_layout(struct inode *ino,
2121		   struct nfs_open_context *ctx,
2122		   loff_t pos,
2123		   u64 count,
2124		   enum pnfs_iomode iomode,
2125		   bool strict_iomode,
2126		   gfp_t gfp_flags)
2127{
2128	struct pnfs_layout_range arg = {
2129		.iomode = iomode,
2130		.offset = pos,
2131		.length = count,
2132	};
2133	unsigned pg_offset;
2134	struct nfs_server *server = NFS_SERVER(ino);
2135	struct nfs_client *clp = server->nfs_client;
2136	struct pnfs_layout_hdr *lo = NULL;
2137	struct pnfs_layout_segment *lseg = NULL;
2138	struct nfs4_layoutget *lgp;
2139	nfs4_stateid stateid;
2140	struct nfs4_exception exception = {
2141		.inode = ino,
2142	};
2143	unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
2144	bool first;
2145
2146	if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
2147		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2148				 PNFS_UPDATE_LAYOUT_NO_PNFS);
2149		goto out;
2150	}
2151
2152	if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
2153		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2154				 PNFS_UPDATE_LAYOUT_MDSTHRESH);
2155		goto out;
2156	}
2157
2158lookup_again:
2159	if (!nfs4_valid_open_stateid(ctx->state)) {
2160		trace_pnfs_update_layout(ino, pos, count,
2161					 iomode, lo, lseg,
2162					 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2163		lseg = ERR_PTR(-EIO);
2164		goto out;
2165	}
2166
2167	lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
2168	if (IS_ERR(lseg))
2169		goto out;
2170	first = false;
2171	spin_lock(&ino->i_lock);
2172	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2173	if (lo == NULL) {
2174		spin_unlock(&ino->i_lock);
2175		lseg = ERR_PTR(-ENOMEM);
2176		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2177				 PNFS_UPDATE_LAYOUT_NOMEM);
2178		goto out;
2179	}
2180
2181	/* Do we even need to bother with this? */
2182	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2183		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2184				 PNFS_UPDATE_LAYOUT_BULK_RECALL);
2185		dprintk("%s matches recall, use MDS\n", __func__);
2186		goto out_unlock;
2187	}
2188
2189	/* if LAYOUTGET already failed once we don't try again */
2190	if (pnfs_layout_io_test_failed(lo, iomode)) {
2191		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2192				 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2193		goto out_unlock;
2194	}
2195
2196	/*
2197	 * If the layout segment list is empty, but there are outstanding
2198	 * layoutget calls, then they might be subject to a layoutrecall.
2199	 */
2200	if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2201	    atomic_read(&lo->plh_outstanding) != 0) {
2202		spin_unlock(&ino->i_lock);
2203		lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2204					   TASK_KILLABLE));
2205		if (IS_ERR(lseg))
2206			goto out_put_layout_hdr;
2207		pnfs_put_layout_hdr(lo);
2208		goto lookup_again;
2209	}
2210
2211	/*
2212	 * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2213	 * for LAYOUTRETURN.
2214	 */
2215	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2216		spin_unlock(&ino->i_lock);
2217		dprintk("%s wait for layoutreturn\n", __func__);
2218		lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2219		if (!IS_ERR(lseg)) {
2220			pnfs_put_layout_hdr(lo);
2221			dprintk("%s retrying\n", __func__);
2222			trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2223						 lseg,
2224						 PNFS_UPDATE_LAYOUT_RETRY);
2225			goto lookup_again;
2226		}
2227		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2228					 PNFS_UPDATE_LAYOUT_RETURN);
2229		goto out_put_layout_hdr;
2230	}
2231
2232	lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2233	if (lseg) {
2234		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2235				PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2236		goto out_unlock;
2237	}
2238
2239	/*
2240	 * Choose a stateid for the LAYOUTGET. If we don't have a layout
2241	 * stateid, or it has been invalidated, then we must use the open
2242	 * stateid.
2243	 */
2244	if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2245		int status;
2246
2247		/*
2248		 * The first layoutget for the file. Need to serialize per
2249		 * RFC 5661 Errata 3208.
2250		 */
2251		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2252				     &lo->plh_flags)) {
2253			spin_unlock(&ino->i_lock);
2254			lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2255						NFS_LAYOUT_FIRST_LAYOUTGET,
2256						TASK_KILLABLE));
2257			if (IS_ERR(lseg))
2258				goto out_put_layout_hdr;
2259			pnfs_put_layout_hdr(lo);
2260			dprintk("%s retrying\n", __func__);
2261			goto lookup_again;
2262		}
2263
2264		spin_unlock(&ino->i_lock);
2265		first = true;
2266		status = nfs4_select_rw_stateid(ctx->state,
2267					iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2268					NULL, &stateid, NULL);
2269		if (status != 0) {
2270			lseg = ERR_PTR(status);
2271			trace_pnfs_update_layout(ino, pos, count,
2272					iomode, lo, lseg,
2273					PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2274			nfs4_schedule_stateid_recovery(server, ctx->state);
2275			pnfs_clear_first_layoutget(lo);
2276			pnfs_put_layout_hdr(lo);
2277			goto lookup_again;
2278		}
2279		spin_lock(&ino->i_lock);
2280	} else {
2281		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2282	}
2283
2284	if (pnfs_layoutgets_blocked(lo)) {
2285		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2286				PNFS_UPDATE_LAYOUT_BLOCKED);
2287		goto out_unlock;
2288	}
2289	nfs_layoutget_begin(lo);
2290	spin_unlock(&ino->i_lock);
2291
2292	_add_to_server_list(lo, server);
2293
2294	pg_offset = arg.offset & ~PAGE_MASK;
2295	if (pg_offset) {
2296		arg.offset -= pg_offset;
2297		arg.length += pg_offset;
2298	}
2299	if (arg.length != NFS4_MAX_UINT64)
2300		arg.length = PAGE_ALIGN(arg.length);
2301
2302	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2303	if (!lgp) {
2304		lseg = ERR_PTR(-ENOMEM);
2305		trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2306					 PNFS_UPDATE_LAYOUT_NOMEM);
2307		nfs_layoutget_end(lo);
2308		goto out_put_layout_hdr;
2309	}
2310
2311	lgp->lo = lo;
2312	pnfs_get_layout_hdr(lo);
2313
2314	lseg = nfs4_proc_layoutget(lgp, &exception);
2315	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2316				 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2317	nfs_layoutget_end(lo);
2318	if (IS_ERR(lseg)) {
2319		switch(PTR_ERR(lseg)) {
2320		case -EBUSY:
2321			if (time_after(jiffies, giveup))
2322				lseg = NULL;
2323			break;
2324		case -ERECALLCONFLICT:
2325		case -EAGAIN:
2326			break;
2327		case -ENODATA:
2328			/* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2329			pnfs_layout_set_fail_bit(
2330				lo, pnfs_iomode_to_fail_bit(iomode));
2331			lseg = NULL;
2332			goto out_put_layout_hdr;
2333		default:
2334			if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2335				pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2336				lseg = NULL;
2337			}
2338			goto out_put_layout_hdr;
2339		}
2340		if (lseg) {
2341			if (!exception.retry)
2342				goto out_put_layout_hdr;
2343			if (first)
2344				pnfs_clear_first_layoutget(lo);
2345			trace_pnfs_update_layout(ino, pos, count,
2346				iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2347			pnfs_put_layout_hdr(lo);
2348			goto lookup_again;
2349		}
2350	} else {
2351		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2352	}
2353
2354out_put_layout_hdr:
2355	if (first)
2356		pnfs_clear_first_layoutget(lo);
2357	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2358				 PNFS_UPDATE_LAYOUT_EXIT);
2359	pnfs_put_layout_hdr(lo);
2360out:
2361	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2362			"(%s, offset: %llu, length: %llu)\n",
2363			__func__, ino->i_sb->s_id,
2364			(unsigned long long)NFS_FILEID(ino),
2365			IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2366			iomode==IOMODE_RW ?  "read/write" : "read-only",
2367			(unsigned long long)pos,
2368			(unsigned long long)count);
2369	return lseg;
2370out_unlock:
2371	spin_unlock(&ino->i_lock);
2372	goto out_put_layout_hdr;
2373}
2374EXPORT_SYMBOL_GPL(pnfs_update_layout);
2375
2376static bool
2377pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2378{
2379	switch (range->iomode) {
2380	case IOMODE_READ:
2381	case IOMODE_RW:
2382		break;
2383	default:
2384		return false;
2385	}
2386	if (range->offset == NFS4_MAX_UINT64)
2387		return false;
2388	if (range->length == 0)
2389		return false;
2390	if (range->length != NFS4_MAX_UINT64 &&
2391	    range->length > NFS4_MAX_UINT64 - range->offset)
2392		return false;
2393	return true;
2394}
2395
2396static struct pnfs_layout_hdr *
2397_pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2398{
2399	struct pnfs_layout_hdr *lo;
2400
2401	spin_lock(&ino->i_lock);
2402	lo = pnfs_find_alloc_layout(ino, ctx, nfs_io_gfp_mask());
2403	if (!lo)
2404		goto out_unlock;
2405	if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2406		goto out_unlock;
2407	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2408		goto out_unlock;
2409	if (pnfs_layoutgets_blocked(lo))
2410		goto out_unlock;
2411	if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2412		goto out_unlock;
2413	nfs_layoutget_begin(lo);
2414	spin_unlock(&ino->i_lock);
2415	_add_to_server_list(lo, NFS_SERVER(ino));
2416	return lo;
2417
2418out_unlock:
2419	spin_unlock(&ino->i_lock);
2420	pnfs_put_layout_hdr(lo);
2421	return NULL;
2422}
2423
2424static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2425				     struct nfs_open_context *ctx)
2426{
2427	struct inode *ino = data->dentry->d_inode;
2428	struct pnfs_layout_range rng = {
2429		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2430			  IOMODE_RW: IOMODE_READ,
2431		.offset = 0,
2432		.length = NFS4_MAX_UINT64,
2433	};
2434	struct nfs4_layoutget *lgp;
2435	struct pnfs_layout_hdr *lo;
2436
2437	/* Heuristic: don't send layoutget if we have cached data */
2438	if (rng.iomode == IOMODE_READ &&
2439	   (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2440		return;
2441
2442	lo = _pnfs_grab_empty_layout(ino, ctx);
2443	if (!lo)
2444		return;
2445	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2446					     nfs_io_gfp_mask());
2447	if (!lgp) {
2448		pnfs_clear_first_layoutget(lo);
2449		nfs_layoutget_end(lo);
2450		pnfs_put_layout_hdr(lo);
2451		return;
2452	}
2453	lgp->lo = lo;
2454	data->lgp = lgp;
2455	data->o_arg.lg_args = &lgp->args;
2456	data->o_res.lg_res = &lgp->res;
2457}
2458
2459static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2460				     struct nfs_open_context *ctx)
2461{
2462	struct inode *ino = data->dentry->d_inode;
2463	struct pnfs_layout_range rng = {
2464		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2465			  IOMODE_RW: IOMODE_READ,
2466		.offset = 0,
2467		.length = NFS4_MAX_UINT64,
2468	};
2469	struct nfs4_layoutget *lgp;
2470
2471	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2472					     nfs_io_gfp_mask());
2473	if (!lgp)
2474		return;
2475	data->lgp = lgp;
2476	data->o_arg.lg_args = &lgp->args;
2477	data->o_res.lg_res = &lgp->res;
2478}
2479
2480void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2481			 struct nfs_open_context *ctx)
2482{
2483	struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2484
2485	if (!(pnfs_enabled_sb(server) &&
2486	      server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2487		return;
2488	/* Could check on max_ops, but currently hardcoded high enough */
2489	if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2490		return;
2491	if (data->lgp)
2492		return;
2493	if (data->state)
2494		_lgopen_prepare_attached(data, ctx);
2495	else
2496		_lgopen_prepare_floating(data, ctx);
2497}
2498
2499void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2500		       struct nfs_open_context *ctx)
2501{
2502	struct pnfs_layout_hdr *lo;
2503	struct pnfs_layout_segment *lseg;
2504	struct nfs_server *srv = NFS_SERVER(ino);
2505	u32 iomode;
2506
2507	if (!lgp)
2508		return;
2509	dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2510	if (lgp->res.status) {
2511		switch (lgp->res.status) {
2512		default:
2513			break;
2514		/*
2515		 * Halt lgopen attempts if the server doesn't recognise
2516		 * the "current stateid" value, the layout type, or the
2517		 * layoutget operation as being valid.
2518		 * Also if it complains about too many ops in the compound
2519		 * or of the request/reply being too big.
2520		 */
2521		case -NFS4ERR_BAD_STATEID:
2522		case -NFS4ERR_NOTSUPP:
2523		case -NFS4ERR_REP_TOO_BIG:
2524		case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2525		case -NFS4ERR_REQ_TOO_BIG:
2526		case -NFS4ERR_TOO_MANY_OPS:
2527		case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2528			srv->caps &= ~NFS_CAP_LGOPEN;
2529		}
2530		return;
2531	}
2532	if (!lgp->lo) {
2533		lo = _pnfs_grab_empty_layout(ino, ctx);
2534		if (!lo)
2535			return;
2536		lgp->lo = lo;
2537	} else
2538		lo = lgp->lo;
2539
2540	lseg = pnfs_layout_process(lgp);
2541	if (!IS_ERR(lseg)) {
2542		iomode = lgp->args.range.iomode;
2543		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2544		pnfs_put_lseg(lseg);
2545	}
2546}
2547
2548void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2549{
2550	if (lgp != NULL) {
2551		if (lgp->lo) {
2552			pnfs_clear_first_layoutget(lgp->lo);
2553			nfs_layoutget_end(lgp->lo);
2554		}
2555		pnfs_layoutget_free(lgp);
2556	}
2557}
2558
2559struct pnfs_layout_segment *
2560pnfs_layout_process(struct nfs4_layoutget *lgp)
2561{
2562	struct pnfs_layout_hdr *lo = lgp->lo;
2563	struct nfs4_layoutget_res *res = &lgp->res;
2564	struct pnfs_layout_segment *lseg;
2565	struct inode *ino = lo->plh_inode;
2566	LIST_HEAD(free_me);
2567
2568	if (!pnfs_sanity_check_layout_range(&res->range))
2569		return ERR_PTR(-EINVAL);
2570
2571	/* Inject layout blob into I/O device driver */
2572	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2573	if (IS_ERR_OR_NULL(lseg)) {
2574		if (!lseg)
2575			lseg = ERR_PTR(-ENOMEM);
2576
2577		dprintk("%s: Could not allocate layout: error %ld\n",
2578		       __func__, PTR_ERR(lseg));
2579		return lseg;
2580	}
2581
2582	pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2583
2584	spin_lock(&ino->i_lock);
2585	if (pnfs_layoutgets_blocked(lo)) {
2586		dprintk("%s forget reply due to state\n", __func__);
2587		goto out_forget;
2588	}
2589
2590	if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2591	    !pnfs_is_first_layoutget(lo))
2592		goto out_forget;
2593
2594	if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2595		/* existing state ID, make sure the sequence number matches. */
2596		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2597			if (!pnfs_layout_is_valid(lo))
2598				lo->plh_barrier = 0;
2599			dprintk("%s forget reply due to sequence\n", __func__);
2600			goto out_forget;
2601		}
2602		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2603	} else if (pnfs_layout_is_valid(lo)) {
2604		/*
2605		 * We got an entirely new state ID.  Mark all segments for the
2606		 * inode invalid, and retry the layoutget
2607		 */
2608		struct pnfs_layout_range range = {
2609			.iomode = IOMODE_ANY,
2610			.length = NFS4_MAX_UINT64,
2611		};
2612		pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2613		goto out_forget;
2614	} else {
2615		/* We have a completely new layout */
2616		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2617	}
2618
2619	pnfs_get_lseg(lseg);
2620	pnfs_layout_insert_lseg(lo, lseg, &free_me);
2621
2622
2623	if (res->return_on_close)
2624		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2625
2626	spin_unlock(&ino->i_lock);
2627	pnfs_free_lseg_list(&free_me);
2628	return lseg;
2629
2630out_forget:
2631	spin_unlock(&ino->i_lock);
2632	lseg->pls_layout = lo;
2633	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2634	return ERR_PTR(-EAGAIN);
2635}
2636
2637/**
2638 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2639 * @lo: pointer to layout header
2640 * @tmp_list: list header to be used with pnfs_free_lseg_list()
2641 * @return_range: describe layout segment ranges to be returned
2642 * @seq: stateid seqid to match
2643 *
2644 * This function is mainly intended for use by layoutrecall. It attempts
2645 * to free the layout segment immediately, or else to mark it for return
2646 * as soon as its reference count drops to zero.
2647 *
2648 * Returns
2649 * - 0: a layoutreturn needs to be scheduled.
2650 * - EBUSY: there are layout segment that are still in use.
2651 * - ENOENT: there are no layout segments that need to be returned.
2652 */
2653int
2654pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2655				struct list_head *tmp_list,
2656				const struct pnfs_layout_range *return_range,
2657				u32 seq)
2658{
2659	struct pnfs_layout_segment *lseg, *next;
2660	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
2661	int remaining = 0;
2662
2663	dprintk("%s:Begin lo %p\n", __func__, lo);
2664
2665	assert_spin_locked(&lo->plh_inode->i_lock);
2666
2667	if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2668		tmp_list = &lo->plh_return_segs;
2669
2670	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2671		if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2672			dprintk("%s: marking lseg %p iomode %d "
2673				"offset %llu length %llu\n", __func__,
2674				lseg, lseg->pls_range.iomode,
2675				lseg->pls_range.offset,
2676				lseg->pls_range.length);
2677			if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2678				tmp_list = &lo->plh_return_segs;
2679			if (mark_lseg_invalid(lseg, tmp_list))
2680				continue;
2681			remaining++;
2682			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2683			pnfs_lseg_cancel_io(server, lseg);
2684		}
2685
2686	if (remaining) {
2687		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2688		return -EBUSY;
2689	}
2690
2691	if (!list_empty(&lo->plh_return_segs)) {
2692		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2693		return 0;
2694	}
2695
2696	return -ENOENT;
2697}
2698
2699static void
2700pnfs_mark_layout_for_return(struct inode *inode,
2701			    const struct pnfs_layout_range *range)
2702{
2703	struct pnfs_layout_hdr *lo;
2704	bool return_now = false;
2705
2706	spin_lock(&inode->i_lock);
2707	lo = NFS_I(inode)->layout;
2708	if (!pnfs_layout_is_valid(lo)) {
2709		spin_unlock(&inode->i_lock);
2710		return;
2711	}
2712	pnfs_set_plh_return_info(lo, range->iomode, 0);
2713	/*
2714	 * mark all matching lsegs so that we are sure to have no live
2715	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2716	 * for how it works.
2717	 */
2718	if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2719		const struct cred *cred;
2720		nfs4_stateid stateid;
2721		enum pnfs_iomode iomode;
2722
2723		return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2724		spin_unlock(&inode->i_lock);
2725		if (return_now)
2726			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
2727					       PNFS_FL_LAYOUTRETURN_ASYNC);
2728	} else {
2729		spin_unlock(&inode->i_lock);
2730		nfs_commit_inode(inode, 0);
2731	}
2732}
2733
2734void pnfs_error_mark_layout_for_return(struct inode *inode,
2735				       struct pnfs_layout_segment *lseg)
2736{
2737	struct pnfs_layout_range range = {
2738		.iomode = lseg->pls_range.iomode,
2739		.offset = 0,
2740		.length = NFS4_MAX_UINT64,
2741	};
2742
2743	pnfs_mark_layout_for_return(inode, &range);
2744}
2745EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2746
2747static bool
2748pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2749{
2750	return pnfs_layout_is_valid(lo) &&
2751		!test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2752		!test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2753}
2754
2755static struct pnfs_layout_segment *
2756pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2757		     const struct pnfs_layout_range *range,
2758		     enum pnfs_iomode iomode)
2759{
2760	struct pnfs_layout_segment *lseg;
2761
2762	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2763		if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2764			continue;
2765		if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2766			continue;
2767		if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2768			continue;
2769		if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2770			return lseg;
2771	}
2772	return NULL;
2773}
2774
2775/* Find open file states whose mode matches that of the range */
2776static bool
2777pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2778				 const struct pnfs_layout_range *range)
2779{
2780	struct list_head *head;
2781	struct nfs_open_context *ctx;
2782	fmode_t mode = 0;
2783
2784	if (!pnfs_layout_can_be_returned(lo) ||
2785	    !pnfs_find_first_lseg(lo, range, range->iomode))
2786		return false;
2787
2788	head = &NFS_I(lo->plh_inode)->open_files;
2789	list_for_each_entry_rcu(ctx, head, list) {
2790		if (ctx->state)
2791			mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2792	}
2793
2794	switch (range->iomode) {
2795	default:
2796		break;
2797	case IOMODE_READ:
2798		mode &= ~FMODE_WRITE;
2799		break;
2800	case IOMODE_RW:
2801		if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2802			mode &= ~FMODE_READ;
2803	}
2804	return mode == 0;
2805}
2806
2807static int pnfs_layout_return_unused_byserver(struct nfs_server *server,
2808					      void *data)
2809{
2810	const struct pnfs_layout_range *range = data;
2811	const struct cred *cred;
2812	struct pnfs_layout_hdr *lo;
2813	struct inode *inode;
2814	nfs4_stateid stateid;
2815	enum pnfs_iomode iomode;
2816
2817restart:
2818	rcu_read_lock();
2819	list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2820		inode = lo->plh_inode;
2821		if (!inode || !pnfs_layout_can_be_returned(lo) ||
2822		    test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2823			continue;
2824		spin_lock(&inode->i_lock);
2825		if (!lo->plh_inode ||
2826		    !pnfs_should_return_unused_layout(lo, range)) {
2827			spin_unlock(&inode->i_lock);
2828			continue;
2829		}
2830		pnfs_get_layout_hdr(lo);
2831		pnfs_set_plh_return_info(lo, range->iomode, 0);
2832		if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
2833						    range, 0) != 0 ||
2834		    !pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode)) {
2835			spin_unlock(&inode->i_lock);
2836			rcu_read_unlock();
2837			pnfs_put_layout_hdr(lo);
2838			cond_resched();
2839			goto restart;
2840		}
2841		spin_unlock(&inode->i_lock);
2842		rcu_read_unlock();
2843		pnfs_send_layoutreturn(lo, &stateid, &cred, iomode,
2844				       PNFS_FL_LAYOUTRETURN_ASYNC);
2845		pnfs_put_layout_hdr(lo);
2846		cond_resched();
2847		goto restart;
2848	}
2849	rcu_read_unlock();
2850	return 0;
2851}
2852
2853void
2854pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2855				 enum pnfs_iomode iomode)
2856{
2857	struct pnfs_layout_range range = {
2858		.iomode = iomode,
2859		.offset = 0,
2860		.length = NFS4_MAX_UINT64,
2861	};
2862
2863	nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2864			&range);
2865}
2866
2867/* Check if we have we have a valid layout but if there isn't an intersection
2868 * between the request and the pgio->pg_lseg, put this pgio->pg_lseg away.
2869 */
2870void
2871pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio,
2872			     struct nfs_page *req)
2873{
2874	if (pgio->pg_lseg == NULL ||
2875	    (test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags) &&
2876	    pnfs_lseg_request_intersecting(pgio->pg_lseg, req)))
2877		return;
2878	pnfs_put_lseg(pgio->pg_lseg);
2879	pgio->pg_lseg = NULL;
2880}
2881EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2882
2883void
2884pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2885{
2886	u64 rd_size;
2887
2888	pnfs_generic_pg_check_layout(pgio, req);
2889	if (pgio->pg_lseg == NULL) {
2890		if (pgio->pg_dreq == NULL)
2891			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2892		else
2893			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq,
2894						      req_offset(req));
2895
2896		pgio->pg_lseg =
2897			pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2898					   req_offset(req), rd_size,
2899					   IOMODE_READ, false,
2900					   nfs_io_gfp_mask());
2901		if (IS_ERR(pgio->pg_lseg)) {
2902			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2903			pgio->pg_lseg = NULL;
2904			return;
2905		}
2906	}
2907	/* If no lseg, fall back to read through mds */
2908	if (pgio->pg_lseg == NULL)
2909		nfs_pageio_reset_read_mds(pgio);
2910
2911}
2912EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2913
2914void
2915pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2916			   struct nfs_page *req, u64 wb_size)
2917{
2918	pnfs_generic_pg_check_layout(pgio, req);
2919	if (pgio->pg_lseg == NULL) {
2920		pgio->pg_lseg =
2921			pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2922					   req_offset(req), wb_size, IOMODE_RW,
2923					   false, nfs_io_gfp_mask());
2924		if (IS_ERR(pgio->pg_lseg)) {
2925			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2926			pgio->pg_lseg = NULL;
2927			return;
2928		}
2929	}
2930	/* If no lseg, fall back to write through mds */
2931	if (pgio->pg_lseg == NULL)
2932		nfs_pageio_reset_write_mds(pgio);
2933}
2934EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2935
2936void
2937pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2938{
2939	if (desc->pg_lseg) {
2940		pnfs_put_lseg(desc->pg_lseg);
2941		desc->pg_lseg = NULL;
2942	}
2943}
2944EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2945
2946/*
2947 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2948 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2949 */
2950size_t
2951pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2952		     struct nfs_page *prev, struct nfs_page *req)
2953{
2954	unsigned int size;
2955	u64 seg_end, req_start, seg_left;
2956
2957	size = nfs_generic_pg_test(pgio, prev, req);
2958	if (!size)
2959		return 0;
2960
2961	/*
2962	 * 'size' contains the number of bytes left in the current page (up
2963	 * to the original size asked for in @req->wb_bytes).
2964	 *
2965	 * Calculate how many bytes are left in the layout segment
2966	 * and if there are less bytes than 'size', return that instead.
2967	 *
2968	 * Please also note that 'end_offset' is actually the offset of the
2969	 * first byte that lies outside the pnfs_layout_range. FIXME?
2970	 *
2971	 */
2972	if (pgio->pg_lseg) {
2973		seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2974				     pgio->pg_lseg->pls_range.length);
2975		req_start = req_offset(req);
2976
2977		/* start of request is past the last byte of this segment */
2978		if (req_start >= seg_end)
2979			return 0;
2980
2981		/* adjust 'size' iff there are fewer bytes left in the
2982		 * segment than what nfs_generic_pg_test returned */
2983		seg_left = seg_end - req_start;
2984		if (seg_left < size)
2985			size = (unsigned int)seg_left;
2986	}
2987
2988	return size;
2989}
2990EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2991
2992int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2993{
2994	struct nfs_pageio_descriptor pgio;
2995
2996	/* Resend all requests through the MDS */
2997	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2998			      hdr->completion_ops);
2999	return nfs_pageio_resend(&pgio, hdr);
3000}
3001EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
3002
3003static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
3004{
3005
3006	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
3007	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
3008	    PNFS_LAYOUTRET_ON_ERROR) {
3009		pnfs_return_layout(hdr->inode);
3010	}
3011	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
3012		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
3013}
3014
3015/*
3016 * Called by non rpc-based layout drivers
3017 */
3018void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
3019{
3020	if (likely(!hdr->pnfs_error)) {
3021		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
3022				hdr->mds_offset + hdr->res.count);
3023		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
3024	}
3025	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
3026	if (unlikely(hdr->pnfs_error))
3027		pnfs_ld_handle_write_error(hdr);
3028	hdr->mds_ops->rpc_release(hdr);
3029}
3030EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
3031
3032static void
3033pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
3034		struct nfs_pgio_header *hdr)
3035{
3036	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3037
3038	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3039		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3040		nfs_pageio_reset_write_mds(desc);
3041		mirror->pg_recoalesce = 1;
3042	}
3043	hdr->completion_ops->completion(hdr);
3044}
3045
3046static enum pnfs_try_status
3047pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
3048			const struct rpc_call_ops *call_ops,
3049			struct pnfs_layout_segment *lseg,
3050			int how)
3051{
3052	struct inode *inode = hdr->inode;
3053	enum pnfs_try_status trypnfs;
3054	struct nfs_server *nfss = NFS_SERVER(inode);
3055
3056	hdr->mds_ops = call_ops;
3057
3058	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
3059		inode->i_ino, hdr->args.count, hdr->args.offset, how);
3060	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
3061	if (trypnfs != PNFS_NOT_ATTEMPTED)
3062		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
3063	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3064	return trypnfs;
3065}
3066
3067static void
3068pnfs_do_write(struct nfs_pageio_descriptor *desc,
3069	      struct nfs_pgio_header *hdr, int how)
3070{
3071	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3072	struct pnfs_layout_segment *lseg = desc->pg_lseg;
3073	enum pnfs_try_status trypnfs;
3074
3075	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
3076	switch (trypnfs) {
3077	case PNFS_NOT_ATTEMPTED:
3078		pnfs_write_through_mds(desc, hdr);
3079		break;
3080	case PNFS_ATTEMPTED:
3081		break;
3082	case PNFS_TRY_AGAIN:
3083		/* cleanup hdr and prepare to redo pnfs */
3084		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3085			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3086			list_splice_init(&hdr->pages, &mirror->pg_list);
3087			mirror->pg_recoalesce = 1;
3088		}
3089		hdr->mds_ops->rpc_release(hdr);
3090	}
3091}
3092
3093static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
3094{
3095	pnfs_put_lseg(hdr->lseg);
3096	nfs_pgio_header_free(hdr);
3097}
3098
3099int
3100pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
3101{
3102	struct nfs_pgio_header *hdr;
3103	int ret;
3104
3105	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3106	if (!hdr) {
3107		desc->pg_error = -ENOMEM;
3108		return desc->pg_error;
3109	}
3110	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
3111
3112	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3113	ret = nfs_generic_pgio(desc, hdr);
3114	if (!ret)
3115		pnfs_do_write(desc, hdr, desc->pg_ioflags);
3116
3117	return ret;
3118}
3119EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
3120
3121int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
3122{
3123	struct nfs_pageio_descriptor pgio;
3124
3125	/* Resend all requests through the MDS */
3126	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
3127	return nfs_pageio_resend(&pgio, hdr);
3128}
3129EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
3130
3131static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
3132{
3133	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
3134	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
3135	    PNFS_LAYOUTRET_ON_ERROR) {
3136		pnfs_return_layout(hdr->inode);
3137	}
3138	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
3139		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
3140}
3141
3142/*
3143 * Called by non rpc-based layout drivers
3144 */
3145void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
3146{
3147	if (likely(!hdr->pnfs_error))
3148		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
3149	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
3150	if (unlikely(hdr->pnfs_error))
3151		pnfs_ld_handle_read_error(hdr);
3152	hdr->mds_ops->rpc_release(hdr);
3153}
3154EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
3155
3156static void
3157pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
3158		struct nfs_pgio_header *hdr)
3159{
3160	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3161
3162	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3163		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3164		nfs_pageio_reset_read_mds(desc);
3165		mirror->pg_recoalesce = 1;
3166	}
3167	hdr->completion_ops->completion(hdr);
3168}
3169
3170/*
3171 * Call the appropriate parallel I/O subsystem read function.
3172 */
3173static enum pnfs_try_status
3174pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
3175		       const struct rpc_call_ops *call_ops,
3176		       struct pnfs_layout_segment *lseg)
3177{
3178	struct inode *inode = hdr->inode;
3179	struct nfs_server *nfss = NFS_SERVER(inode);
3180	enum pnfs_try_status trypnfs;
3181
3182	hdr->mds_ops = call_ops;
3183
3184	dprintk("%s: Reading ino:%lu %u@%llu\n",
3185		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3186
3187	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3188	if (trypnfs != PNFS_NOT_ATTEMPTED)
3189		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3190	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3191	return trypnfs;
3192}
3193
3194/* Resend all requests through pnfs. */
3195void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3196			   unsigned int mirror_idx)
3197{
3198	struct nfs_pageio_descriptor pgio;
3199
3200	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3201		/* Prevent deadlocks with layoutreturn! */
3202		pnfs_put_lseg(hdr->lseg);
3203		hdr->lseg = NULL;
3204
3205		nfs_pageio_init_read(&pgio, hdr->inode, false,
3206					hdr->completion_ops);
3207		pgio.pg_mirror_idx = mirror_idx;
3208		hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3209	}
3210}
3211EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3212
3213static void
3214pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3215{
3216	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3217	struct pnfs_layout_segment *lseg = desc->pg_lseg;
3218	enum pnfs_try_status trypnfs;
3219
3220	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3221	switch (trypnfs) {
3222	case PNFS_NOT_ATTEMPTED:
3223		pnfs_read_through_mds(desc, hdr);
3224		break;
3225	case PNFS_ATTEMPTED:
3226		break;
3227	case PNFS_TRY_AGAIN:
3228		/* cleanup hdr and prepare to redo pnfs */
3229		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3230			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3231			list_splice_init(&hdr->pages, &mirror->pg_list);
3232			mirror->pg_recoalesce = 1;
3233		}
3234		hdr->mds_ops->rpc_release(hdr);
3235	}
3236}
3237
3238static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3239{
3240	pnfs_put_lseg(hdr->lseg);
3241	nfs_pgio_header_free(hdr);
3242}
3243
3244int
3245pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3246{
3247	struct nfs_pgio_header *hdr;
3248	int ret;
3249
3250	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3251	if (!hdr) {
3252		desc->pg_error = -ENOMEM;
3253		return desc->pg_error;
3254	}
3255	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3256	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3257	ret = nfs_generic_pgio(desc, hdr);
3258	if (!ret)
3259		pnfs_do_read(desc, hdr);
3260	return ret;
3261}
3262EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3263
3264static void pnfs_clear_layoutcommitting(struct inode *inode)
3265{
3266	unsigned long *bitlock = &NFS_I(inode)->flags;
3267
3268	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3269	smp_mb__after_atomic();
3270	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3271}
3272
3273/*
3274 * There can be multiple RW segments.
3275 */
3276static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3277{
3278	struct pnfs_layout_segment *lseg;
3279
3280	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3281		if (lseg->pls_range.iomode == IOMODE_RW &&
3282		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3283			list_add(&lseg->pls_lc_list, listp);
3284	}
3285}
3286
3287static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3288{
3289	struct pnfs_layout_segment *lseg, *tmp;
3290
3291	/* Matched by references in pnfs_set_layoutcommit */
3292	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3293		list_del_init(&lseg->pls_lc_list);
3294		pnfs_put_lseg(lseg);
3295	}
3296
3297	pnfs_clear_layoutcommitting(inode);
3298}
3299
3300void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3301{
3302	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3303}
3304EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3305
3306void
3307pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3308		loff_t end_pos)
3309{
3310	struct nfs_inode *nfsi = NFS_I(inode);
3311	bool mark_as_dirty = false;
3312
3313	spin_lock(&inode->i_lock);
3314	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3315		nfsi->layout->plh_lwb = end_pos;
3316		mark_as_dirty = true;
3317		dprintk("%s: Set layoutcommit for inode %lu ",
3318			__func__, inode->i_ino);
3319	} else if (end_pos > nfsi->layout->plh_lwb)
3320		nfsi->layout->plh_lwb = end_pos;
3321	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3322		/* references matched in nfs4_layoutcommit_release */
3323		pnfs_get_lseg(lseg);
3324	}
3325	spin_unlock(&inode->i_lock);
3326	dprintk("%s: lseg %p end_pos %llu\n",
3327		__func__, lseg, nfsi->layout->plh_lwb);
3328
3329	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3330	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3331	if (mark_as_dirty)
3332		mark_inode_dirty_sync(inode);
3333}
3334EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3335
3336void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3337{
3338	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3339
3340	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3341		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3342	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3343}
3344
3345/*
3346 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3347 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3348 * data to disk to allow the server to recover the data if it crashes.
3349 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3350 * is off, and a COMMIT is sent to a data server, or
3351 * if WRITEs to a data server return NFS_DATA_SYNC.
3352 */
3353int
3354pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3355{
3356	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3357	struct nfs4_layoutcommit_data *data;
3358	struct nfs_inode *nfsi = NFS_I(inode);
3359	loff_t end_pos;
3360	int status;
3361	bool mark_as_dirty = false;
3362
3363	if (!pnfs_layoutcommit_outstanding(inode))
3364		return 0;
3365
3366	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3367
3368	status = -EAGAIN;
3369	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3370		if (!sync)
3371			goto out;
3372		status = wait_on_bit_lock_action(&nfsi->flags,
3373				NFS_INO_LAYOUTCOMMITTING,
3374				nfs_wait_bit_killable,
3375				TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
3376		if (status)
3377			goto out;
3378	}
3379
3380	status = -ENOMEM;
3381	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3382	data = kzalloc(sizeof(*data), nfs_io_gfp_mask());
3383	if (!data)
3384		goto clear_layoutcommitting;
3385
3386	status = 0;
3387	spin_lock(&inode->i_lock);
3388	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3389		goto out_unlock;
3390
3391	INIT_LIST_HEAD(&data->lseg_list);
3392	pnfs_list_write_lseg(inode, &data->lseg_list);
3393
3394	end_pos = nfsi->layout->plh_lwb;
3395
3396	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3397	data->cred = get_cred(nfsi->layout->plh_lc_cred);
3398	spin_unlock(&inode->i_lock);
3399
3400	data->args.inode = inode;
3401	nfs_fattr_init(&data->fattr);
3402	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3403	data->res.fattr = &data->fattr;
3404	if (end_pos != 0)
3405		data->args.lastbytewritten = end_pos - 1;
3406	else
3407		data->args.lastbytewritten = U64_MAX;
3408	data->res.server = NFS_SERVER(inode);
3409
3410	if (ld->prepare_layoutcommit) {
3411		status = ld->prepare_layoutcommit(&data->args);
3412		if (status) {
3413			if (status != -ENOSPC)
3414				put_cred(data->cred);
3415			spin_lock(&inode->i_lock);
3416			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3417			if (end_pos > nfsi->layout->plh_lwb)
3418				nfsi->layout->plh_lwb = end_pos;
3419			if (status != -ENOSPC)
3420				goto out_unlock;
3421			spin_unlock(&inode->i_lock);
3422			mark_as_dirty = true;
3423		}
3424	}
3425
3426
3427	status = nfs4_proc_layoutcommit(data, sync);
3428out:
3429	if (status || mark_as_dirty)
3430		mark_inode_dirty_sync(inode);
3431	dprintk("<-- %s status %d\n", __func__, status);
3432	return status;
3433out_unlock:
3434	spin_unlock(&inode->i_lock);
3435	kfree(data);
3436clear_layoutcommitting:
3437	pnfs_clear_layoutcommitting(inode);
3438	goto out;
3439}
3440EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3441
3442int
3443pnfs_generic_sync(struct inode *inode, bool datasync)
3444{
3445	return pnfs_layoutcommit_inode(inode, true);
3446}
3447EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3448
3449struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3450{
3451	struct nfs4_threshold *thp;
3452
3453	thp = kzalloc(sizeof(*thp), nfs_io_gfp_mask());
3454	if (!thp) {
3455		dprintk("%s mdsthreshold allocation failed\n", __func__);
3456		return NULL;
3457	}
3458	return thp;
3459}
3460
3461#if IS_ENABLED(CONFIG_NFS_V4_2)
3462int
3463pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3464{
3465	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3466	struct nfs_server *server = NFS_SERVER(inode);
3467	struct nfs_inode *nfsi = NFS_I(inode);
3468	struct nfs42_layoutstat_data *data;
3469	struct pnfs_layout_hdr *hdr;
3470	int status = 0;
3471
3472	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3473		goto out;
3474
3475	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3476		goto out;
3477
3478	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3479		goto out;
3480
3481	spin_lock(&inode->i_lock);
3482	if (!NFS_I(inode)->layout) {
3483		spin_unlock(&inode->i_lock);
3484		goto out_clear_layoutstats;
3485	}
3486	hdr = NFS_I(inode)->layout;
3487	pnfs_get_layout_hdr(hdr);
3488	spin_unlock(&inode->i_lock);
3489
3490	data = kzalloc(sizeof(*data), gfp_flags);
3491	if (!data) {
3492		status = -ENOMEM;
3493		goto out_put;
3494	}
3495
3496	data->args.fh = NFS_FH(inode);
3497	data->args.inode = inode;
3498	status = ld->prepare_layoutstats(&data->args);
3499	if (status)
3500		goto out_free;
3501
3502	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3503
3504out:
3505	dprintk("%s returns %d\n", __func__, status);
3506	return status;
3507
3508out_free:
3509	kfree(data);
3510out_put:
3511	pnfs_put_layout_hdr(hdr);
3512out_clear_layoutstats:
3513	smp_mb__before_atomic();
3514	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3515	smp_mb__after_atomic();
3516	goto out;
3517}
3518EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3519#endif
3520
3521unsigned int layoutstats_timer;
3522module_param(layoutstats_timer, uint, 0644);
3523EXPORT_SYMBOL_GPL(layoutstats_timer);