Linux Audio

Check our new training course

Embedded Linux training

Mar 31-Apr 8, 2025
Register
Loading...
v4.17
   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
  41#define NFSDBG_FACILITY		NFSDBG_PNFS
  42#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  43
  44/* Locking:
  45 *
  46 * pnfs_spinlock:
  47 *      protects pnfs_modules_tbl.
  48 */
  49static DEFINE_SPINLOCK(pnfs_spinlock);
  50
  51/*
  52 * pnfs_modules_tbl holds all pnfs modules
  53 */
  54static LIST_HEAD(pnfs_modules_tbl);
  55
  56static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  57static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
  58		struct list_head *free_me,
  59		const struct pnfs_layout_range *range,
  60		u32 seq);
  61static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
  62		                struct list_head *tmp_list);
  63
  64/* Return the registered pnfs layout driver module matching given id */
  65static struct pnfs_layoutdriver_type *
  66find_pnfs_driver_locked(u32 id)
  67{
  68	struct pnfs_layoutdriver_type *local;
  69
  70	list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  71		if (local->id == id)
  72			goto out;
  73	local = NULL;
  74out:
  75	dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  76	return local;
  77}
  78
  79static struct pnfs_layoutdriver_type *
  80find_pnfs_driver(u32 id)
  81{
  82	struct pnfs_layoutdriver_type *local;
  83
  84	spin_lock(&pnfs_spinlock);
  85	local = find_pnfs_driver_locked(id);
  86	if (local != NULL && !try_module_get(local->owner)) {
  87		dprintk("%s: Could not grab reference on module\n", __func__);
  88		local = NULL;
  89	}
  90	spin_unlock(&pnfs_spinlock);
  91	return local;
  92}
  93
  94void
  95unset_pnfs_layoutdriver(struct nfs_server *nfss)
  96{
  97	if (nfss->pnfs_curr_ld) {
  98		if (nfss->pnfs_curr_ld->clear_layoutdriver)
  99			nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 100		/* Decrement the MDS count. Purge the deviceid cache if zero */
 101		if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 102			nfs4_deviceid_purge_client(nfss->nfs_client);
 103		module_put(nfss->pnfs_curr_ld->owner);
 104	}
 105	nfss->pnfs_curr_ld = NULL;
 106}
 107
 108/*
 109 * When the server sends a list of layout types, we choose one in the order
 110 * given in the list below.
 111 *
 112 * FIXME: should this list be configurable in some fashion? module param?
 113 * 	  mount option? something else?
 114 */
 115static const u32 ld_prefs[] = {
 116	LAYOUT_SCSI,
 117	LAYOUT_BLOCK_VOLUME,
 118	LAYOUT_OSD2_OBJECTS,
 119	LAYOUT_FLEX_FILES,
 120	LAYOUT_NFSV4_1_FILES,
 121	0
 122};
 123
 124static int
 125ld_cmp(const void *e1, const void *e2)
 126{
 127	u32 ld1 = *((u32 *)e1);
 128	u32 ld2 = *((u32 *)e2);
 129	int i;
 130
 131	for (i = 0; ld_prefs[i] != 0; i++) {
 132		if (ld1 == ld_prefs[i])
 133			return -1;
 134
 135		if (ld2 == ld_prefs[i])
 136			return 1;
 137	}
 138	return 0;
 139}
 140
 141/*
 142 * Try to set the server's pnfs module to the pnfs layout type specified by id.
 143 * Currently only one pNFS layout driver per filesystem is supported.
 144 *
 145 * @ids array of layout types supported by MDS.
 146 */
 147void
 148set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 149		      struct nfs_fsinfo *fsinfo)
 150{
 151	struct pnfs_layoutdriver_type *ld_type = NULL;
 152	u32 id;
 153	int i;
 154
 155	if (fsinfo->nlayouttypes == 0)
 156		goto out_no_driver;
 157	if (!(server->nfs_client->cl_exchange_flags &
 158		 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 159		printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 160			__func__, server->nfs_client->cl_exchange_flags);
 161		goto out_no_driver;
 162	}
 163
 164	sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 165		sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 166
 167	for (i = 0; i < fsinfo->nlayouttypes; i++) {
 168		id = fsinfo->layouttype[i];
 169		ld_type = find_pnfs_driver(id);
 170		if (!ld_type) {
 171			request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 172					id);
 173			ld_type = find_pnfs_driver(id);
 174		}
 175		if (ld_type)
 176			break;
 177	}
 178
 179	if (!ld_type) {
 180		dprintk("%s: No pNFS module found!\n", __func__);
 181		goto out_no_driver;
 182	}
 183
 184	server->pnfs_curr_ld = ld_type;
 185	if (ld_type->set_layoutdriver
 186	    && ld_type->set_layoutdriver(server, mntfh)) {
 187		printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 188			"driver %u.\n", __func__, id);
 189		module_put(ld_type->owner);
 190		goto out_no_driver;
 191	}
 192	/* Bump the MDS count */
 193	atomic_inc(&server->nfs_client->cl_mds_count);
 194
 195	dprintk("%s: pNFS module for %u set\n", __func__, id);
 196	return;
 197
 198out_no_driver:
 199	dprintk("%s: Using NFSv4 I/O\n", __func__);
 200	server->pnfs_curr_ld = NULL;
 201}
 202
 203int
 204pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 205{
 206	int status = -EINVAL;
 207	struct pnfs_layoutdriver_type *tmp;
 208
 209	if (ld_type->id == 0) {
 210		printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 211		return status;
 212	}
 213	if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 214		printk(KERN_ERR "NFS: %s Layout driver must provide "
 215		       "alloc_lseg and free_lseg.\n", __func__);
 216		return status;
 217	}
 218
 219	spin_lock(&pnfs_spinlock);
 220	tmp = find_pnfs_driver_locked(ld_type->id);
 221	if (!tmp) {
 222		list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 223		status = 0;
 224		dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 225			ld_type->name);
 226	} else {
 227		printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 228			__func__, ld_type->id);
 229	}
 230	spin_unlock(&pnfs_spinlock);
 231
 232	return status;
 233}
 234EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 235
 236void
 237pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 238{
 239	dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 240	spin_lock(&pnfs_spinlock);
 241	list_del(&ld_type->pnfs_tblid);
 242	spin_unlock(&pnfs_spinlock);
 243}
 244EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 245
 246/*
 247 * pNFS client layout cache
 248 */
 249
 250/* Need to hold i_lock if caller does not already hold reference */
 251void
 252pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 253{
 254	refcount_inc(&lo->plh_refcount);
 255}
 256
 257static struct pnfs_layout_hdr *
 258pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 259{
 260	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 261	return ld->alloc_layout_hdr(ino, gfp_flags);
 262}
 263
 264static void
 265pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 266{
 267	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 268	struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 269
 270	if (!list_empty(&lo->plh_layouts)) {
 271		struct nfs_client *clp = server->nfs_client;
 272
 273		spin_lock(&clp->cl_lock);
 274		list_del_init(&lo->plh_layouts);
 275		spin_unlock(&clp->cl_lock);
 276	}
 277	put_rpccred(lo->plh_lc_cred);
 278	return ld->free_layout_hdr(lo);
 279}
 280
 281static void
 282pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 283{
 284	struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 285	dprintk("%s: freeing layout cache %p\n", __func__, lo);
 286	nfsi->layout = NULL;
 287	/* Reset MDS Threshold I/O counters */
 288	nfsi->write_io = 0;
 289	nfsi->read_io = 0;
 290}
 291
 292void
 293pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 294{
 295	struct inode *inode;
 296
 297	if (!lo)
 298		return;
 299	inode = lo->plh_inode;
 300	pnfs_layoutreturn_before_put_layout_hdr(lo);
 301
 302	if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 303		if (!list_empty(&lo->plh_segs))
 304			WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 305		pnfs_detach_layout_hdr(lo);
 306		spin_unlock(&inode->i_lock);
 307		pnfs_free_layout_hdr(lo);
 308	}
 309}
 310
 311static void
 312pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 313			 u32 seq)
 314{
 315	if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 316		iomode = IOMODE_ANY;
 317	lo->plh_return_iomode = iomode;
 318	set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 319	if (seq != 0) {
 320		WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 321		lo->plh_return_seq = seq;
 322	}
 323}
 324
 325static void
 326pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 327{
 328	struct pnfs_layout_segment *lseg;
 329	lo->plh_return_iomode = 0;
 330	lo->plh_return_seq = 0;
 331	clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 332	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 333		if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 334			continue;
 335		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 336	}
 337}
 338
 339static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 340{
 341	clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 342	clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 343	smp_mb__after_atomic();
 344	wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 345	rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 346}
 347
 348static void
 349pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 350		struct list_head *free_me)
 351{
 352	clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 353	clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 354	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 355		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 356	if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 357		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 358}
 359
 360/*
 361 * Update the seqid of a layout stateid
 
 362 */
 363bool nfs4_refresh_layout_stateid(nfs4_stateid *dst, struct inode *inode)
 
 
 364{
 365	struct pnfs_layout_hdr *lo;
 
 
 
 
 
 366	bool ret = false;
 
 
 367
 368	spin_lock(&inode->i_lock);
 369	lo = NFS_I(inode)->layout;
 370	if (lo && nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 371		dst->seqid = lo->plh_stateid.seqid;
 372		ret = true;
 
 
 
 
 
 
 
 
 
 
 
 
 373	}
 
 374	spin_unlock(&inode->i_lock);
 
 375	return ret;
 376}
 377
 378/*
 379 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 380 *
 381 * In order to continue using the pnfs_layout_hdr, a full recovery
 382 * is required.
 383 * Note that caller must hold inode->i_lock.
 384 */
 385int
 386pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 387		struct list_head *lseg_list)
 388{
 389	struct pnfs_layout_range range = {
 390		.iomode = IOMODE_ANY,
 391		.offset = 0,
 392		.length = NFS4_MAX_UINT64,
 393	};
 394	struct pnfs_layout_segment *lseg, *next;
 395
 396	set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 397	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 398		pnfs_clear_lseg_state(lseg, lseg_list);
 399	pnfs_clear_layoutreturn_info(lo);
 400	pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 401	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 402	    !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 403		pnfs_clear_layoutreturn_waitbit(lo);
 404	return !list_empty(&lo->plh_segs);
 405}
 406
 407static int
 408pnfs_iomode_to_fail_bit(u32 iomode)
 409{
 410	return iomode == IOMODE_RW ?
 411		NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 412}
 413
 414static void
 415pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 416{
 417	lo->plh_retry_timestamp = jiffies;
 418	if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 419		refcount_inc(&lo->plh_refcount);
 420}
 421
 422static void
 423pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 424{
 425	if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 426		refcount_dec(&lo->plh_refcount);
 427}
 428
 429static void
 430pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 431{
 432	struct inode *inode = lo->plh_inode;
 433	struct pnfs_layout_range range = {
 434		.iomode = iomode,
 435		.offset = 0,
 436		.length = NFS4_MAX_UINT64,
 437	};
 438	LIST_HEAD(head);
 439
 440	spin_lock(&inode->i_lock);
 441	pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 442	pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 443	spin_unlock(&inode->i_lock);
 444	pnfs_free_lseg_list(&head);
 445	dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 446			iomode == IOMODE_RW ?  "RW" : "READ");
 447}
 448
 449static bool
 450pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 451{
 452	unsigned long start, end;
 453	int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 454
 455	if (test_bit(fail_bit, &lo->plh_flags) == 0)
 456		return false;
 457	end = jiffies;
 458	start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 459	if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 460		/* It is time to retry the failed layoutgets */
 461		pnfs_layout_clear_fail_bit(lo, fail_bit);
 462		return false;
 463	}
 464	return true;
 465}
 466
 467static void
 468pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 469		const struct pnfs_layout_range *range,
 470		const nfs4_stateid *stateid)
 471{
 472	INIT_LIST_HEAD(&lseg->pls_list);
 473	INIT_LIST_HEAD(&lseg->pls_lc_list);
 474	refcount_set(&lseg->pls_refcount, 1);
 475	set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 476	lseg->pls_layout = lo;
 477	lseg->pls_range = *range;
 478	lseg->pls_seq = be32_to_cpu(stateid->seqid);
 479}
 480
 481static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 482{
 483	if (lseg != NULL) {
 484		struct inode *inode = lseg->pls_layout->plh_inode;
 485		NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 486	}
 487}
 488
 489static void
 490pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 491		struct pnfs_layout_segment *lseg)
 492{
 493	WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 494	list_del_init(&lseg->pls_list);
 495	/* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 496	refcount_dec(&lo->plh_refcount);
 497	if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 498		return;
 499	if (list_empty(&lo->plh_segs) &&
 500	    !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 501	    !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 502		if (atomic_read(&lo->plh_outstanding) == 0)
 503			set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 504		clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 505	}
 506}
 507
 508static bool
 509pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 510		struct pnfs_layout_segment *lseg)
 511{
 512	if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 513	    pnfs_layout_is_valid(lo)) {
 514		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 515		list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 516		return true;
 517	}
 518	return false;
 519}
 520
 521void
 522pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 523{
 524	struct pnfs_layout_hdr *lo;
 525	struct inode *inode;
 526
 527	if (!lseg)
 528		return;
 529
 530	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 531		refcount_read(&lseg->pls_refcount),
 532		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 533
 534	lo = lseg->pls_layout;
 535	inode = lo->plh_inode;
 536
 537	if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 538		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 539			spin_unlock(&inode->i_lock);
 540			return;
 541		}
 542		pnfs_get_layout_hdr(lo);
 543		pnfs_layout_remove_lseg(lo, lseg);
 544		if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 545			lseg = NULL;
 546		spin_unlock(&inode->i_lock);
 547		pnfs_free_lseg(lseg);
 548		pnfs_put_layout_hdr(lo);
 549	}
 550}
 551EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 552
 553/*
 554 * is l2 fully contained in l1?
 555 *   start1                             end1
 556 *   [----------------------------------)
 557 *           start2           end2
 558 *           [----------------)
 559 */
 560static bool
 561pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 562		 const struct pnfs_layout_range *l2)
 563{
 564	u64 start1 = l1->offset;
 565	u64 end1 = pnfs_end_offset(start1, l1->length);
 566	u64 start2 = l2->offset;
 567	u64 end2 = pnfs_end_offset(start2, l2->length);
 568
 569	return (start1 <= start2) && (end1 >= end2);
 570}
 571
 572static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 573		struct list_head *tmp_list)
 574{
 575	if (!refcount_dec_and_test(&lseg->pls_refcount))
 576		return false;
 577	pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 578	list_add(&lseg->pls_list, tmp_list);
 579	return true;
 580}
 581
 582/* Returns 1 if lseg is removed from list, 0 otherwise */
 583static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 584			     struct list_head *tmp_list)
 585{
 586	int rv = 0;
 587
 588	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 589		/* Remove the reference keeping the lseg in the
 590		 * list.  It will now be removed when all
 591		 * outstanding io is finished.
 592		 */
 593		dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 594			refcount_read(&lseg->pls_refcount));
 595		if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 596			rv = 1;
 597	}
 598	return rv;
 599}
 600
 601/*
 602 * Compare 2 layout stateid sequence ids, to see which is newer,
 603 * taking into account wraparound issues.
 604 */
 605static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 606{
 607	return (s32)(s1 - s2) > 0;
 608}
 609
 610static bool
 611pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 612		 const struct pnfs_layout_range *recall_range)
 613{
 614	return (recall_range->iomode == IOMODE_ANY ||
 615		lseg_range->iomode == recall_range->iomode) &&
 616	       pnfs_lseg_range_intersecting(lseg_range, recall_range);
 617}
 618
 619static bool
 620pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 621		const struct pnfs_layout_range *recall_range,
 622		u32 seq)
 623{
 624	if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 625		return false;
 626	if (recall_range == NULL)
 627		return true;
 628	return pnfs_should_free_range(&lseg->pls_range, recall_range);
 629}
 630
 631/**
 632 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 633 * @lo: layout header containing the lsegs
 634 * @tmp_list: list head where doomed lsegs should go
 635 * @recall_range: optional recall range argument to match (may be NULL)
 636 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 637 *
 638 * Walk the list of lsegs in the layout header, and tear down any that should
 639 * be destroyed. If "recall_range" is specified then the segment must match
 640 * that range. If "seq" is non-zero, then only match segments that were handed
 641 * out at or before that sequence.
 642 *
 643 * Returns number of matching invalid lsegs remaining in list after scanning
 644 * it and purging them.
 645 */
 646int
 647pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 648			    struct list_head *tmp_list,
 649			    const struct pnfs_layout_range *recall_range,
 650			    u32 seq)
 651{
 652	struct pnfs_layout_segment *lseg, *next;
 653	int remaining = 0;
 654
 655	dprintk("%s:Begin lo %p\n", __func__, lo);
 656
 657	if (list_empty(&lo->plh_segs))
 658		return 0;
 659	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 660		if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 661			dprintk("%s: freeing lseg %p iomode %d seq %u "
 662				"offset %llu length %llu\n", __func__,
 663				lseg, lseg->pls_range.iomode, lseg->pls_seq,
 664				lseg->pls_range.offset, lseg->pls_range.length);
 665			if (!mark_lseg_invalid(lseg, tmp_list))
 666				remaining++;
 667		}
 668	dprintk("%s:Return %i\n", __func__, remaining);
 669	return remaining;
 670}
 671
 672static void
 673pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 674		struct list_head *free_me,
 675		const struct pnfs_layout_range *range,
 676		u32 seq)
 677{
 678	struct pnfs_layout_segment *lseg, *next;
 679
 680	list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 681		if (pnfs_match_lseg_recall(lseg, range, seq))
 682			list_move_tail(&lseg->pls_list, free_me);
 683	}
 684}
 685
 686/* note free_me must contain lsegs from a single layout_hdr */
 687void
 688pnfs_free_lseg_list(struct list_head *free_me)
 689{
 690	struct pnfs_layout_segment *lseg, *tmp;
 691
 692	if (list_empty(free_me))
 693		return;
 694
 695	list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 696		list_del(&lseg->pls_list);
 697		pnfs_free_lseg(lseg);
 698	}
 699}
 700
 701void
 702pnfs_destroy_layout(struct nfs_inode *nfsi)
 703{
 704	struct pnfs_layout_hdr *lo;
 705	LIST_HEAD(tmp_list);
 706
 707	spin_lock(&nfsi->vfs_inode.i_lock);
 708	lo = nfsi->layout;
 709	if (lo) {
 710		pnfs_get_layout_hdr(lo);
 711		pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 712		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 713		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 714		spin_unlock(&nfsi->vfs_inode.i_lock);
 715		pnfs_free_lseg_list(&tmp_list);
 716		nfs_commit_inode(&nfsi->vfs_inode, 0);
 717		pnfs_put_layout_hdr(lo);
 718	} else
 719		spin_unlock(&nfsi->vfs_inode.i_lock);
 720}
 721EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 722
 723static bool
 724pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 725		struct list_head *layout_list)
 726{
 727	struct pnfs_layout_hdr *lo;
 728	bool ret = false;
 729
 730	spin_lock(&inode->i_lock);
 731	lo = NFS_I(inode)->layout;
 732	if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 733		pnfs_get_layout_hdr(lo);
 734		list_add(&lo->plh_bulk_destroy, layout_list);
 735		ret = true;
 736	}
 737	spin_unlock(&inode->i_lock);
 738	return ret;
 739}
 740
 741/* Caller must hold rcu_read_lock and clp->cl_lock */
 742static int
 743pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 744		struct nfs_server *server,
 745		struct list_head *layout_list)
 
 
 746{
 747	struct pnfs_layout_hdr *lo, *next;
 748	struct inode *inode;
 749
 750	list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 751		if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
 
 
 752			continue;
 
 
 
 753		inode = igrab(lo->plh_inode);
 754		if (inode == NULL)
 755			continue;
 756		list_del_init(&lo->plh_layouts);
 757		if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 758			continue;
 759		rcu_read_unlock();
 760		spin_unlock(&clp->cl_lock);
 761		iput(inode);
 
 
 
 
 
 
 762		spin_lock(&clp->cl_lock);
 763		rcu_read_lock();
 764		return -EAGAIN;
 765	}
 766	return 0;
 767}
 768
 769static int
 770pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 771		bool is_bulk_recall)
 772{
 773	struct pnfs_layout_hdr *lo;
 774	struct inode *inode;
 775	LIST_HEAD(lseg_list);
 776	int ret = 0;
 777
 778	while (!list_empty(layout_list)) {
 779		lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 780				plh_bulk_destroy);
 781		dprintk("%s freeing layout for inode %lu\n", __func__,
 782			lo->plh_inode->i_ino);
 783		inode = lo->plh_inode;
 784
 785		pnfs_layoutcommit_inode(inode, false);
 786
 787		spin_lock(&inode->i_lock);
 788		list_del_init(&lo->plh_bulk_destroy);
 789		if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 790			if (is_bulk_recall)
 791				set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 792			ret = -EAGAIN;
 793		}
 794		spin_unlock(&inode->i_lock);
 795		pnfs_free_lseg_list(&lseg_list);
 796		/* Free all lsegs that are attached to commit buckets */
 797		nfs_commit_inode(inode, 0);
 798		pnfs_put_layout_hdr(lo);
 799		iput(inode);
 800	}
 801	return ret;
 802}
 803
 804int
 805pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 806		struct nfs_fsid *fsid,
 807		bool is_recall)
 808{
 809	struct nfs_server *server;
 810	LIST_HEAD(layout_list);
 811
 812	spin_lock(&clp->cl_lock);
 813	rcu_read_lock();
 814restart:
 815	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 816		if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 817			continue;
 818		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 819				server,
 820				&layout_list) != 0)
 821			goto restart;
 822	}
 823	rcu_read_unlock();
 824	spin_unlock(&clp->cl_lock);
 825
 826	if (list_empty(&layout_list))
 827		return 0;
 828	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 829}
 830
 831int
 832pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 833		bool is_recall)
 834{
 835	struct nfs_server *server;
 836	LIST_HEAD(layout_list);
 837
 838	spin_lock(&clp->cl_lock);
 839	rcu_read_lock();
 840restart:
 841	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 842		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 843					server,
 844					&layout_list) != 0)
 845			goto restart;
 846	}
 847	rcu_read_unlock();
 848	spin_unlock(&clp->cl_lock);
 849
 850	if (list_empty(&layout_list))
 851		return 0;
 852	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 853}
 854
 855/*
 856 * Called by the state manger to remove all layouts established under an
 857 * expired lease.
 858 */
 859void
 860pnfs_destroy_all_layouts(struct nfs_client *clp)
 861{
 862	nfs4_deviceid_mark_client_invalid(clp);
 863	nfs4_deviceid_purge_client(clp);
 864
 865	pnfs_destroy_layouts_byclid(clp, false);
 866}
 867
 868/* update lo->plh_stateid with new if is more recent */
 869void
 870pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 871			bool update_barrier)
 872{
 873	u32 oldseq, newseq, new_barrier = 0;
 874
 875	oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 876	newseq = be32_to_cpu(new->seqid);
 877
 878	if (!pnfs_layout_is_valid(lo)) {
 879		nfs4_stateid_copy(&lo->plh_stateid, new);
 880		lo->plh_barrier = newseq;
 881		pnfs_clear_layoutreturn_info(lo);
 882		clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 883		return;
 884	}
 885	if (pnfs_seqid_is_newer(newseq, oldseq)) {
 886		nfs4_stateid_copy(&lo->plh_stateid, new);
 887		/*
 888		 * Because of wraparound, we want to keep the barrier
 889		 * "close" to the current seqids.
 890		 */
 891		new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 892	}
 893	if (update_barrier)
 894		new_barrier = be32_to_cpu(new->seqid);
 895	else if (new_barrier == 0)
 896		return;
 897	if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 898		lo->plh_barrier = new_barrier;
 899}
 900
 901static bool
 902pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 903		const nfs4_stateid *stateid)
 904{
 905	u32 seqid = be32_to_cpu(stateid->seqid);
 906
 907	return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 908}
 909
 910/* lget is set to 1 if called from inside send_layoutget call chain */
 911static bool
 912pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 913{
 914	return lo->plh_block_lgets ||
 915		test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 916}
 917
 918/*
 919 * Get layout from server.
 920 *    for now, assume that whole file layouts are requested.
 921 *    arg->offset: 0
 922 *    arg->length: all ones
 923 */
 924static struct pnfs_layout_segment *
 925send_layoutget(struct pnfs_layout_hdr *lo,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 926	   struct nfs_open_context *ctx,
 927	   nfs4_stateid *stateid,
 928	   const struct pnfs_layout_range *range,
 929	   long *timeout, gfp_t gfp_flags)
 930{
 931	struct inode *ino = lo->plh_inode;
 932	struct nfs_server *server = NFS_SERVER(ino);
 
 933	struct nfs4_layoutget *lgp;
 934	loff_t i_size;
 935
 936	dprintk("--> %s\n", __func__);
 937
 938	/*
 939	 * Synchronously retrieve layout information from server and
 940	 * store in lseg. If we race with a concurrent seqid morphing
 941	 * op, then re-send the LAYOUTGET.
 942	 */
 943	lgp = kzalloc(sizeof(*lgp), gfp_flags);
 944	if (lgp == NULL)
 945		return ERR_PTR(-ENOMEM);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 946
 947	i_size = i_size_read(ino);
 
 948
 949	lgp->args.minlength = PAGE_SIZE;
 950	if (lgp->args.minlength > range->length)
 951		lgp->args.minlength = range->length;
 952	if (range->iomode == IOMODE_READ) {
 953		if (range->offset >= i_size)
 954			lgp->args.minlength = 0;
 955		else if (i_size - range->offset < lgp->args.minlength)
 956			lgp->args.minlength = i_size - range->offset;
 
 
 
 
 957	}
 958	lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 959	pnfs_copy_range(&lgp->args.range, range);
 960	lgp->args.type = server->pnfs_curr_ld->id;
 961	lgp->args.inode = ino;
 962	lgp->args.ctx = get_nfs_open_context(ctx);
 963	nfs4_stateid_copy(&lgp->args.stateid, stateid);
 964	lgp->gfp_flags = gfp_flags;
 965	lgp->cred = lo->plh_lc_cred;
 
 
 
 
 
 
 966
 967	return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
 
 
 
 
 
 968}
 969
 970static void pnfs_clear_layoutcommit(struct inode *inode,
 971		struct list_head *head)
 972{
 973	struct nfs_inode *nfsi = NFS_I(inode);
 974	struct pnfs_layout_segment *lseg, *tmp;
 975
 976	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 977		return;
 978	list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 979		if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 980			continue;
 981		pnfs_lseg_dec_and_remove_zero(lseg, head);
 982	}
 983}
 984
 985void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
 986		const nfs4_stateid *arg_stateid,
 987		const struct pnfs_layout_range *range,
 988		const nfs4_stateid *stateid)
 989{
 990	struct inode *inode = lo->plh_inode;
 991	LIST_HEAD(freeme);
 992
 993	spin_lock(&inode->i_lock);
 994	if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
 995	    !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
 996		goto out_unlock;
 997	if (stateid) {
 998		u32 seq = be32_to_cpu(arg_stateid->seqid);
 999
1000		pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1001		pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1002		pnfs_set_layout_stateid(lo, stateid, true);
1003	} else
1004		pnfs_mark_layout_stateid_invalid(lo, &freeme);
1005out_unlock:
1006	pnfs_clear_layoutreturn_waitbit(lo);
1007	spin_unlock(&inode->i_lock);
1008	pnfs_free_lseg_list(&freeme);
1009
1010}
1011
1012static bool
1013pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1014		nfs4_stateid *stateid,
1015		enum pnfs_iomode *iomode)
1016{
1017	/* Serialise LAYOUTGET/LAYOUTRETURN */
1018	if (atomic_read(&lo->plh_outstanding) != 0)
1019		return false;
1020	if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1021		return false;
1022	set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1023	pnfs_get_layout_hdr(lo);
1024	if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1025		if (stateid != NULL) {
1026			nfs4_stateid_copy(stateid, &lo->plh_stateid);
1027			if (lo->plh_return_seq != 0)
1028				stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1029		}
1030		if (iomode != NULL)
1031			*iomode = lo->plh_return_iomode;
1032		pnfs_clear_layoutreturn_info(lo);
1033		return true;
1034	}
1035	if (stateid != NULL)
1036		nfs4_stateid_copy(stateid, &lo->plh_stateid);
1037	if (iomode != NULL)
1038		*iomode = IOMODE_ANY;
1039	return true;
1040}
1041
1042static void
1043pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1044		struct pnfs_layout_hdr *lo,
1045		const nfs4_stateid *stateid,
1046		enum pnfs_iomode iomode)
1047{
1048	struct inode *inode = lo->plh_inode;
1049
1050	args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1051	args->inode = inode;
1052	args->range.iomode = iomode;
1053	args->range.offset = 0;
1054	args->range.length = NFS4_MAX_UINT64;
1055	args->layout = lo;
1056	nfs4_stateid_copy(&args->stateid, stateid);
1057}
1058
1059static int
1060pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1061		       enum pnfs_iomode iomode, bool sync)
1062{
1063	struct inode *ino = lo->plh_inode;
1064	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1065	struct nfs4_layoutreturn *lrp;
1066	int status = 0;
1067
1068	lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1069	if (unlikely(lrp == NULL)) {
1070		status = -ENOMEM;
1071		spin_lock(&ino->i_lock);
1072		pnfs_clear_layoutreturn_waitbit(lo);
1073		spin_unlock(&ino->i_lock);
1074		pnfs_put_layout_hdr(lo);
1075		goto out;
1076	}
1077
1078	pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1079	lrp->args.ld_private = &lrp->ld_private;
1080	lrp->clp = NFS_SERVER(ino)->nfs_client;
1081	lrp->cred = lo->plh_lc_cred;
1082	if (ld->prepare_layoutreturn)
1083		ld->prepare_layoutreturn(&lrp->args);
1084
1085	status = nfs4_proc_layoutreturn(lrp, sync);
1086out:
1087	dprintk("<-- %s status: %d\n", __func__, status);
1088	return status;
1089}
1090
1091/* Return true if layoutreturn is needed */
1092static bool
1093pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1094{
1095	struct pnfs_layout_segment *s;
 
 
1096
1097	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1098		return false;
1099
1100	/* Defer layoutreturn until all lsegs are done */
 
 
 
1101	list_for_each_entry(s, &lo->plh_segs, pls_list) {
 
 
 
 
1102		if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1103			return false;
1104	}
1105
1106	return true;
1107}
1108
1109static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1110{
1111	struct inode *inode= lo->plh_inode;
1112
1113	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1114		return;
1115	spin_lock(&inode->i_lock);
1116	if (pnfs_layout_need_return(lo)) {
1117		nfs4_stateid stateid;
1118		enum pnfs_iomode iomode;
1119		bool send;
1120
1121		send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1122		spin_unlock(&inode->i_lock);
1123		if (send) {
1124			/* Send an async layoutreturn so we dont deadlock */
1125			pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1126		}
1127	} else
1128		spin_unlock(&inode->i_lock);
1129}
1130
1131/*
1132 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1133 * when the layout segment list is empty.
1134 *
1135 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1136 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1137 * deviceid is marked invalid.
1138 */
1139int
1140_pnfs_return_layout(struct inode *ino)
1141{
1142	struct pnfs_layout_hdr *lo = NULL;
1143	struct nfs_inode *nfsi = NFS_I(ino);
1144	LIST_HEAD(tmp_list);
1145	nfs4_stateid stateid;
1146	int status = 0;
1147	bool send;
1148
1149	dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1150
1151	spin_lock(&ino->i_lock);
1152	lo = nfsi->layout;
1153	if (!lo) {
1154		spin_unlock(&ino->i_lock);
1155		dprintk("NFS: %s no layout to return\n", __func__);
1156		goto out;
1157	}
1158	/* Reference matched in nfs4_layoutreturn_release */
1159	pnfs_get_layout_hdr(lo);
1160	/* Is there an outstanding layoutreturn ? */
1161	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1162		spin_unlock(&ino->i_lock);
1163		if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1164					TASK_UNINTERRUPTIBLE))
1165			goto out_put_layout_hdr;
1166		spin_lock(&ino->i_lock);
1167	}
 
1168	pnfs_clear_layoutcommit(ino, &tmp_list);
1169	pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1170
1171	if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1172		struct pnfs_layout_range range = {
1173			.iomode		= IOMODE_ANY,
1174			.offset		= 0,
1175			.length		= NFS4_MAX_UINT64,
1176		};
1177		NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1178	}
1179
1180	/* Don't send a LAYOUTRETURN if list was initially empty */
1181	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
 
1182		spin_unlock(&ino->i_lock);
1183		dprintk("NFS: %s no layout segments to return\n", __func__);
1184		goto out_put_layout_hdr;
1185	}
1186
1187	send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1188	spin_unlock(&ino->i_lock);
1189	if (send)
1190		status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1191out_put_layout_hdr:
1192	pnfs_free_lseg_list(&tmp_list);
1193	pnfs_put_layout_hdr(lo);
1194out:
1195	dprintk("<-- %s status: %d\n", __func__, status);
1196	return status;
1197}
1198
1199int
1200pnfs_commit_and_return_layout(struct inode *inode)
1201{
1202	struct pnfs_layout_hdr *lo;
1203	int ret;
1204
1205	spin_lock(&inode->i_lock);
1206	lo = NFS_I(inode)->layout;
1207	if (lo == NULL) {
1208		spin_unlock(&inode->i_lock);
1209		return 0;
1210	}
1211	pnfs_get_layout_hdr(lo);
1212	/* Block new layoutgets and read/write to ds */
1213	lo->plh_block_lgets++;
1214	spin_unlock(&inode->i_lock);
1215	filemap_fdatawait(inode->i_mapping);
1216	ret = pnfs_layoutcommit_inode(inode, true);
1217	if (ret == 0)
1218		ret = _pnfs_return_layout(inode);
1219	spin_lock(&inode->i_lock);
1220	lo->plh_block_lgets--;
1221	spin_unlock(&inode->i_lock);
1222	pnfs_put_layout_hdr(lo);
1223	return ret;
1224}
1225
1226bool pnfs_roc(struct inode *ino,
1227		struct nfs4_layoutreturn_args *args,
1228		struct nfs4_layoutreturn_res *res,
1229		const struct rpc_cred *cred)
1230{
1231	struct nfs_inode *nfsi = NFS_I(ino);
1232	struct nfs_open_context *ctx;
1233	struct nfs4_state *state;
1234	struct pnfs_layout_hdr *lo;
1235	struct pnfs_layout_segment *lseg, *next;
1236	nfs4_stateid stateid;
1237	enum pnfs_iomode iomode = 0;
1238	bool layoutreturn = false, roc = false;
1239	bool skip_read = false;
1240
1241	if (!nfs_have_layout(ino))
1242		return false;
1243retry:
 
1244	spin_lock(&ino->i_lock);
1245	lo = nfsi->layout;
1246	if (!lo || !pnfs_layout_is_valid(lo) ||
1247	    test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1248		lo = NULL;
1249		goto out_noroc;
1250	}
1251	pnfs_get_layout_hdr(lo);
1252	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1253		spin_unlock(&ino->i_lock);
 
1254		wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1255				TASK_UNINTERRUPTIBLE);
1256		pnfs_put_layout_hdr(lo);
1257		goto retry;
1258	}
1259
1260	/* no roc if we hold a delegation */
1261	if (nfs4_check_delegation(ino, FMODE_READ)) {
1262		if (nfs4_check_delegation(ino, FMODE_WRITE))
1263			goto out_noroc;
1264		skip_read = true;
1265	}
1266
1267	list_for_each_entry(ctx, &nfsi->open_files, list) {
1268		state = ctx->state;
1269		if (state == NULL)
1270			continue;
1271		/* Don't return layout if there is open file state */
1272		if (state->state & FMODE_WRITE)
1273			goto out_noroc;
1274		if (state->state & FMODE_READ)
1275			skip_read = true;
1276	}
1277
1278
1279	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1280		if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1281			continue;
1282		/* If we are sending layoutreturn, invalidate all valid lsegs */
1283		if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1284			continue;
1285		/*
1286		 * Note: mark lseg for return so pnfs_layout_remove_lseg
1287		 * doesn't invalidate the layout for us.
1288		 */
1289		set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1290		if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1291			continue;
1292		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1293	}
1294
1295	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1296		goto out_noroc;
1297
1298	/* ROC in two conditions:
1299	 * 1. there are ROC lsegs
1300	 * 2. we don't send layoutreturn
1301	 */
1302	/* lo ref dropped in pnfs_roc_release() */
1303	layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1304	/* If the creds don't match, we can't compound the layoutreturn */
1305	if (!layoutreturn || cred != lo->plh_lc_cred)
1306		goto out_noroc;
1307
1308	roc = layoutreturn;
1309	pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1310	res->lrs_present = 0;
1311	layoutreturn = false;
1312
1313out_noroc:
1314	spin_unlock(&ino->i_lock);
 
1315	pnfs_layoutcommit_inode(ino, true);
1316	if (roc) {
1317		struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1318		if (ld->prepare_layoutreturn)
1319			ld->prepare_layoutreturn(args);
1320		pnfs_put_layout_hdr(lo);
1321		return true;
1322	}
1323	if (layoutreturn)
1324		pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1325	pnfs_put_layout_hdr(lo);
1326	return false;
1327}
1328
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1329void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1330		struct nfs4_layoutreturn_res *res,
1331		int ret)
1332{
1333	struct pnfs_layout_hdr *lo = args->layout;
1334	const nfs4_stateid *arg_stateid = NULL;
1335	const nfs4_stateid *res_stateid = NULL;
1336	struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1337
1338	if (ret == 0) {
1339		arg_stateid = &args->stateid;
 
 
1340		if (res->lrs_present)
1341			res_stateid = &res->stateid;
 
 
 
1342	}
1343	pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1344			res_stateid);
1345	if (ld_private && ld_private->ops && ld_private->ops->free)
1346		ld_private->ops->free(ld_private);
1347	pnfs_put_layout_hdr(lo);
1348	trace_nfs4_layoutreturn_on_close(args->inode, 0);
1349}
1350
1351bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1352{
1353	struct nfs_inode *nfsi = NFS_I(ino);
1354        struct pnfs_layout_hdr *lo;
1355        bool sleep = false;
1356
1357	/* we might not have grabbed lo reference. so need to check under
1358	 * i_lock */
1359        spin_lock(&ino->i_lock);
1360        lo = nfsi->layout;
1361        if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1362                rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1363                sleep = true;
1364	}
1365        spin_unlock(&ino->i_lock);
1366        return sleep;
1367}
1368
1369/*
1370 * Compare two layout segments for sorting into layout cache.
1371 * We want to preferentially return RW over RO layouts, so ensure those
1372 * are seen first.
1373 */
1374static s64
1375pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1376	   const struct pnfs_layout_range *l2)
1377{
1378	s64 d;
1379
1380	/* high offset > low offset */
1381	d = l1->offset - l2->offset;
1382	if (d)
1383		return d;
1384
1385	/* short length > long length */
1386	d = l2->length - l1->length;
1387	if (d)
1388		return d;
1389
1390	/* read > read/write */
1391	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1392}
1393
1394static bool
1395pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1396		const struct pnfs_layout_range *l2)
1397{
1398	return pnfs_lseg_range_cmp(l1, l2) > 0;
1399}
1400
1401static bool
1402pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1403		struct pnfs_layout_segment *old)
1404{
1405	return false;
1406}
1407
1408void
1409pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1410		   struct pnfs_layout_segment *lseg,
1411		   bool (*is_after)(const struct pnfs_layout_range *,
1412			   const struct pnfs_layout_range *),
1413		   bool (*do_merge)(struct pnfs_layout_segment *,
1414			   struct pnfs_layout_segment *),
1415		   struct list_head *free_me)
1416{
1417	struct pnfs_layout_segment *lp, *tmp;
1418
1419	dprintk("%s:Begin\n", __func__);
1420
1421	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1422		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1423			continue;
1424		if (do_merge(lseg, lp)) {
1425			mark_lseg_invalid(lp, free_me);
1426			continue;
1427		}
1428		if (is_after(&lseg->pls_range, &lp->pls_range))
1429			continue;
1430		list_add_tail(&lseg->pls_list, &lp->pls_list);
1431		dprintk("%s: inserted lseg %p "
1432			"iomode %d offset %llu length %llu before "
1433			"lp %p iomode %d offset %llu length %llu\n",
1434			__func__, lseg, lseg->pls_range.iomode,
1435			lseg->pls_range.offset, lseg->pls_range.length,
1436			lp, lp->pls_range.iomode, lp->pls_range.offset,
1437			lp->pls_range.length);
1438		goto out;
1439	}
1440	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1441	dprintk("%s: inserted lseg %p "
1442		"iomode %d offset %llu length %llu at tail\n",
1443		__func__, lseg, lseg->pls_range.iomode,
1444		lseg->pls_range.offset, lseg->pls_range.length);
1445out:
1446	pnfs_get_layout_hdr(lo);
1447
1448	dprintk("%s:Return\n", __func__);
1449}
1450EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1451
1452static void
1453pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1454		   struct pnfs_layout_segment *lseg,
1455		   struct list_head *free_me)
1456{
1457	struct inode *inode = lo->plh_inode;
1458	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1459
1460	if (ld->add_lseg != NULL)
1461		ld->add_lseg(lo, lseg, free_me);
1462	else
1463		pnfs_generic_layout_insert_lseg(lo, lseg,
1464				pnfs_lseg_range_is_after,
1465				pnfs_lseg_no_merge,
1466				free_me);
1467}
1468
1469static struct pnfs_layout_hdr *
1470alloc_init_layout_hdr(struct inode *ino,
1471		      struct nfs_open_context *ctx,
1472		      gfp_t gfp_flags)
1473{
1474	struct pnfs_layout_hdr *lo;
1475
1476	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1477	if (!lo)
1478		return NULL;
1479	refcount_set(&lo->plh_refcount, 1);
1480	INIT_LIST_HEAD(&lo->plh_layouts);
1481	INIT_LIST_HEAD(&lo->plh_segs);
1482	INIT_LIST_HEAD(&lo->plh_return_segs);
1483	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1484	lo->plh_inode = ino;
1485	lo->plh_lc_cred = get_rpccred(ctx->cred);
1486	lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1487	return lo;
1488}
1489
1490static struct pnfs_layout_hdr *
1491pnfs_find_alloc_layout(struct inode *ino,
1492		       struct nfs_open_context *ctx,
1493		       gfp_t gfp_flags)
1494	__releases(&ino->i_lock)
1495	__acquires(&ino->i_lock)
1496{
1497	struct nfs_inode *nfsi = NFS_I(ino);
1498	struct pnfs_layout_hdr *new = NULL;
1499
1500	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1501
1502	if (nfsi->layout != NULL)
1503		goto out_existing;
1504	spin_unlock(&ino->i_lock);
1505	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1506	spin_lock(&ino->i_lock);
1507
1508	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1509		nfsi->layout = new;
1510		return new;
1511	} else if (new != NULL)
1512		pnfs_free_layout_hdr(new);
1513out_existing:
1514	pnfs_get_layout_hdr(nfsi->layout);
1515	return nfsi->layout;
1516}
1517
1518/*
1519 * iomode matching rules:
1520 * iomode	lseg	strict match
1521 *                      iomode
1522 * -----	-----	------ -----
1523 * ANY		READ	N/A    true
1524 * ANY		RW	N/A    true
1525 * RW		READ	N/A    false
1526 * RW		RW	N/A    true
1527 * READ		READ	N/A    true
1528 * READ		RW	true   false
1529 * READ		RW	false  true
1530 */
1531static bool
1532pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1533		 const struct pnfs_layout_range *range,
1534		 bool strict_iomode)
1535{
1536	struct pnfs_layout_range range1;
1537
1538	if ((range->iomode == IOMODE_RW &&
1539	     ls_range->iomode != IOMODE_RW) ||
1540	    (range->iomode != ls_range->iomode &&
1541	     strict_iomode) ||
1542	    !pnfs_lseg_range_intersecting(ls_range, range))
1543		return 0;
1544
1545	/* range1 covers only the first byte in the range */
1546	range1 = *range;
1547	range1.length = 1;
1548	return pnfs_lseg_range_contained(ls_range, &range1);
1549}
1550
1551/*
1552 * lookup range in layout
1553 */
1554static struct pnfs_layout_segment *
1555pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1556		struct pnfs_layout_range *range,
1557		bool strict_iomode)
1558{
1559	struct pnfs_layout_segment *lseg, *ret = NULL;
1560
1561	dprintk("%s:Begin\n", __func__);
1562
1563	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1564		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1565		    !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1566		    pnfs_lseg_range_match(&lseg->pls_range, range,
1567					  strict_iomode)) {
1568			ret = pnfs_get_lseg(lseg);
1569			break;
1570		}
1571	}
1572
1573	dprintk("%s:Return lseg %p ref %d\n",
1574		__func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1575	return ret;
1576}
1577
1578/*
1579 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1580 * to the MDS or over pNFS
1581 *
1582 * The nfs_inode read_io and write_io fields are cumulative counters reset
1583 * when there are no layout segments. Note that in pnfs_update_layout iomode
1584 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1585 * WRITE request.
1586 *
1587 * A return of true means use MDS I/O.
1588 *
1589 * From rfc 5661:
1590 * If a file's size is smaller than the file size threshold, data accesses
1591 * SHOULD be sent to the metadata server.  If an I/O request has a length that
1592 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1593 * server.  If both file size and I/O size are provided, the client SHOULD
1594 * reach or exceed  both thresholds before sending its read or write
1595 * requests to the data server.
1596 */
1597static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1598				     struct inode *ino, int iomode)
1599{
1600	struct nfs4_threshold *t = ctx->mdsthreshold;
1601	struct nfs_inode *nfsi = NFS_I(ino);
1602	loff_t fsize = i_size_read(ino);
1603	bool size = false, size_set = false, io = false, io_set = false, ret = false;
1604
1605	if (t == NULL)
1606		return ret;
1607
1608	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1609		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1610
1611	switch (iomode) {
1612	case IOMODE_READ:
1613		if (t->bm & THRESHOLD_RD) {
1614			dprintk("%s fsize %llu\n", __func__, fsize);
1615			size_set = true;
1616			if (fsize < t->rd_sz)
1617				size = true;
1618		}
1619		if (t->bm & THRESHOLD_RD_IO) {
1620			dprintk("%s nfsi->read_io %llu\n", __func__,
1621				nfsi->read_io);
1622			io_set = true;
1623			if (nfsi->read_io < t->rd_io_sz)
1624				io = true;
1625		}
1626		break;
1627	case IOMODE_RW:
1628		if (t->bm & THRESHOLD_WR) {
1629			dprintk("%s fsize %llu\n", __func__, fsize);
1630			size_set = true;
1631			if (fsize < t->wr_sz)
1632				size = true;
1633		}
1634		if (t->bm & THRESHOLD_WR_IO) {
1635			dprintk("%s nfsi->write_io %llu\n", __func__,
1636				nfsi->write_io);
1637			io_set = true;
1638			if (nfsi->write_io < t->wr_io_sz)
1639				io = true;
1640		}
1641		break;
1642	}
1643	if (size_set && io_set) {
1644		if (size && io)
1645			ret = true;
1646	} else if (size || io)
1647		ret = true;
1648
1649	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1650	return ret;
1651}
1652
1653static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1654{
1655	/*
1656	 * send layoutcommit as it can hold up layoutreturn due to lseg
1657	 * reference
1658	 */
1659	pnfs_layoutcommit_inode(lo->plh_inode, false);
1660	return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1661				   nfs_wait_bit_killable,
1662				   TASK_UNINTERRUPTIBLE);
 
 
 
 
 
 
 
 
 
 
 
1663}
1664
1665static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1666{
1667	unsigned long *bitlock = &lo->plh_flags;
1668
1669	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1670	smp_mb__after_atomic();
1671	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1672}
1673
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1674/*
1675 * Layout segment is retreived from the server if not cached.
1676 * The appropriate layout segment is referenced and returned to the caller.
1677 */
1678struct pnfs_layout_segment *
1679pnfs_update_layout(struct inode *ino,
1680		   struct nfs_open_context *ctx,
1681		   loff_t pos,
1682		   u64 count,
1683		   enum pnfs_iomode iomode,
1684		   bool strict_iomode,
1685		   gfp_t gfp_flags)
1686{
1687	struct pnfs_layout_range arg = {
1688		.iomode = iomode,
1689		.offset = pos,
1690		.length = count,
1691	};
1692	unsigned pg_offset;
1693	struct nfs_server *server = NFS_SERVER(ino);
1694	struct nfs_client *clp = server->nfs_client;
1695	struct pnfs_layout_hdr *lo = NULL;
1696	struct pnfs_layout_segment *lseg = NULL;
 
1697	nfs4_stateid stateid;
1698	long timeout = 0;
1699	unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1700	bool first;
1701
1702	if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1703		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1704				 PNFS_UPDATE_LAYOUT_NO_PNFS);
1705		goto out;
1706	}
1707
1708	if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1709		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1710				 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1711		goto out;
1712	}
1713
1714	if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1715		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1716				 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1717		goto out;
1718	}
1719
1720lookup_again:
1721	nfs4_client_recover_expired_lease(clp);
 
 
1722	first = false;
1723	spin_lock(&ino->i_lock);
1724	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1725	if (lo == NULL) {
1726		spin_unlock(&ino->i_lock);
1727		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1728				 PNFS_UPDATE_LAYOUT_NOMEM);
1729		goto out;
1730	}
1731
1732	/* Do we even need to bother with this? */
1733	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1734		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1735				 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1736		dprintk("%s matches recall, use MDS\n", __func__);
1737		goto out_unlock;
1738	}
1739
1740	/* if LAYOUTGET already failed once we don't try again */
1741	if (pnfs_layout_io_test_failed(lo, iomode)) {
1742		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1743				 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1744		goto out_unlock;
1745	}
1746
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1747	lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1748	if (lseg) {
1749		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1750				PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1751		goto out_unlock;
1752	}
1753
1754	if (!nfs4_valid_open_stateid(ctx->state)) {
1755		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1756				PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1757		goto out_unlock;
1758	}
1759
1760	/*
1761	 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1762	 * stateid, or it has been invalidated, then we must use the open
1763	 * stateid.
1764	 */
1765	if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 
1766
1767		/*
1768		 * The first layoutget for the file. Need to serialize per
1769		 * RFC 5661 Errata 3208.
1770		 */
1771		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1772				     &lo->plh_flags)) {
1773			spin_unlock(&ino->i_lock);
1774			wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1775				    TASK_UNINTERRUPTIBLE);
 
 
 
1776			pnfs_put_layout_hdr(lo);
1777			dprintk("%s retrying\n", __func__);
1778			goto lookup_again;
1779		}
1780
1781		first = true;
1782		if (nfs4_select_rw_stateid(ctx->state,
1783					iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
1784					NULL, &stateid, NULL) != 0) {
 
 
1785			trace_pnfs_update_layout(ino, pos, count,
1786					iomode, lo, lseg,
1787					PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1788			goto out_unlock;
 
 
 
 
 
 
1789		}
1790	} else {
1791		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1792	}
1793
1794	/*
1795	 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1796	 * for LAYOUTRETURN even if first is true.
1797	 */
1798	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1799		spin_unlock(&ino->i_lock);
1800		dprintk("%s wait for layoutreturn\n", __func__);
1801		if (pnfs_prepare_to_retry_layoutget(lo)) {
 
1802			if (first)
1803				pnfs_clear_first_layoutget(lo);
1804			pnfs_put_layout_hdr(lo);
1805			dprintk("%s retrying\n", __func__);
1806			trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1807					lseg, PNFS_UPDATE_LAYOUT_RETRY);
1808			goto lookup_again;
1809		}
1810		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1811				PNFS_UPDATE_LAYOUT_RETURN);
1812		goto out_put_layout_hdr;
1813	}
1814
1815	if (pnfs_layoutgets_blocked(lo)) {
1816		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1817				PNFS_UPDATE_LAYOUT_BLOCKED);
1818		goto out_unlock;
1819	}
1820	atomic_inc(&lo->plh_outstanding);
1821	spin_unlock(&ino->i_lock);
1822
1823	if (list_empty(&lo->plh_layouts)) {
1824		/* The lo must be on the clp list if there is any
1825		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1826		 */
1827		spin_lock(&clp->cl_lock);
1828		if (list_empty(&lo->plh_layouts))
1829			list_add_tail(&lo->plh_layouts, &server->layouts);
1830		spin_unlock(&clp->cl_lock);
1831	}
1832
1833	pg_offset = arg.offset & ~PAGE_MASK;
1834	if (pg_offset) {
1835		arg.offset -= pg_offset;
1836		arg.length += pg_offset;
1837	}
1838	if (arg.length != NFS4_MAX_UINT64)
1839		arg.length = PAGE_ALIGN(arg.length);
1840
1841	lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
 
 
 
 
 
 
 
 
1842	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1843				 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1844	atomic_dec(&lo->plh_outstanding);
1845	if (IS_ERR(lseg)) {
1846		switch(PTR_ERR(lseg)) {
1847		case -EBUSY:
1848			if (time_after(jiffies, giveup))
1849				lseg = NULL;
1850			break;
1851		case -ERECALLCONFLICT:
1852			/* Huh? We hold no layouts, how is there a recall? */
1853			if (first) {
1854				lseg = NULL;
1855				break;
1856			}
1857			/* Destroy the existing layout and start over */
1858			if (time_after(jiffies, giveup))
1859				pnfs_destroy_layout(NFS_I(ino));
1860			/* Fallthrough */
1861		case -EAGAIN:
1862			break;
1863		default:
1864			if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1865				pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1866				lseg = NULL;
1867			}
1868			goto out_put_layout_hdr;
1869		}
1870		if (lseg) {
1871			if (first)
1872				pnfs_clear_first_layoutget(lo);
1873			trace_pnfs_update_layout(ino, pos, count,
1874				iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1875			pnfs_put_layout_hdr(lo);
1876			goto lookup_again;
1877		}
1878	} else {
1879		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1880	}
1881
1882out_put_layout_hdr:
1883	if (first)
1884		pnfs_clear_first_layoutget(lo);
 
 
1885	pnfs_put_layout_hdr(lo);
1886out:
1887	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1888			"(%s, offset: %llu, length: %llu)\n",
1889			__func__, ino->i_sb->s_id,
1890			(unsigned long long)NFS_FILEID(ino),
1891			IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1892			iomode==IOMODE_RW ?  "read/write" : "read-only",
1893			(unsigned long long)pos,
1894			(unsigned long long)count);
1895	return lseg;
1896out_unlock:
1897	spin_unlock(&ino->i_lock);
1898	goto out_put_layout_hdr;
1899}
1900EXPORT_SYMBOL_GPL(pnfs_update_layout);
1901
1902static bool
1903pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1904{
1905	switch (range->iomode) {
1906	case IOMODE_READ:
1907	case IOMODE_RW:
1908		break;
1909	default:
1910		return false;
1911	}
1912	if (range->offset == NFS4_MAX_UINT64)
1913		return false;
1914	if (range->length == 0)
1915		return false;
1916	if (range->length != NFS4_MAX_UINT64 &&
1917	    range->length > NFS4_MAX_UINT64 - range->offset)
1918		return false;
1919	return true;
1920}
1921
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1922struct pnfs_layout_segment *
1923pnfs_layout_process(struct nfs4_layoutget *lgp)
1924{
1925	struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1926	struct nfs4_layoutget_res *res = &lgp->res;
1927	struct pnfs_layout_segment *lseg;
1928	struct inode *ino = lo->plh_inode;
1929	LIST_HEAD(free_me);
1930
1931	if (!pnfs_sanity_check_layout_range(&res->range))
1932		return ERR_PTR(-EINVAL);
1933
1934	/* Inject layout blob into I/O device driver */
1935	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1936	if (IS_ERR_OR_NULL(lseg)) {
1937		if (!lseg)
1938			lseg = ERR_PTR(-ENOMEM);
1939
1940		dprintk("%s: Could not allocate layout: error %ld\n",
1941		       __func__, PTR_ERR(lseg));
1942		return lseg;
1943	}
1944
1945	pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1946
1947	spin_lock(&ino->i_lock);
1948	if (pnfs_layoutgets_blocked(lo)) {
1949		dprintk("%s forget reply due to state\n", __func__);
1950		goto out_forget;
1951	}
1952
1953	if (!pnfs_layout_is_valid(lo)) {
1954		/* We have a completely new layout */
1955		pnfs_set_layout_stateid(lo, &res->stateid, true);
1956	} else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1957		/* existing state ID, make sure the sequence number matches. */
1958		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1959			dprintk("%s forget reply due to sequence\n", __func__);
1960			goto out_forget;
1961		}
1962		pnfs_set_layout_stateid(lo, &res->stateid, false);
1963	} else {
1964		/*
1965		 * We got an entirely new state ID.  Mark all segments for the
1966		 * inode invalid, and retry the layoutget
1967		 */
1968		pnfs_mark_layout_stateid_invalid(lo, &free_me);
1969		goto out_forget;
1970	}
1971
1972	pnfs_get_lseg(lseg);
1973	pnfs_layout_insert_lseg(lo, lseg, &free_me);
1974
1975
1976	if (res->return_on_close)
1977		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1978
1979	spin_unlock(&ino->i_lock);
1980	pnfs_free_lseg_list(&free_me);
1981	return lseg;
1982
1983out_forget:
1984	spin_unlock(&ino->i_lock);
1985	lseg->pls_layout = lo;
1986	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1987	if (!pnfs_layout_is_valid(lo))
1988		nfs_commit_inode(ino, 0);
1989	return ERR_PTR(-EAGAIN);
1990}
1991
 
 
 
 
 
 
 
 
 
 
1992/**
1993 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1994 * @lo: pointer to layout header
1995 * @tmp_list: list header to be used with pnfs_free_lseg_list()
1996 * @return_range: describe layout segment ranges to be returned
 
1997 *
1998 * This function is mainly intended for use by layoutrecall. It attempts
1999 * to free the layout segment immediately, or else to mark it for return
2000 * as soon as its reference count drops to zero.
 
 
 
 
 
2001 */
2002int
2003pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2004				struct list_head *tmp_list,
2005				const struct pnfs_layout_range *return_range,
2006				u32 seq)
2007{
2008	struct pnfs_layout_segment *lseg, *next;
2009	int remaining = 0;
2010
2011	dprintk("%s:Begin lo %p\n", __func__, lo);
2012
2013	if (list_empty(&lo->plh_segs))
2014		return 0;
2015
2016	assert_spin_locked(&lo->plh_inode->i_lock);
2017
2018	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2019		if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2020			dprintk("%s: marking lseg %p iomode %d "
2021				"offset %llu length %llu\n", __func__,
2022				lseg, lseg->pls_range.iomode,
2023				lseg->pls_range.offset,
2024				lseg->pls_range.length);
2025			if (mark_lseg_invalid(lseg, tmp_list))
2026				continue;
2027			remaining++;
2028			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2029		}
2030
2031	if (remaining)
2032		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
 
 
2033
2034	return remaining;
 
 
 
 
 
2035}
2036
2037void pnfs_error_mark_layout_for_return(struct inode *inode,
2038				       struct pnfs_layout_segment *lseg)
2039{
2040	struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2041	struct pnfs_layout_range range = {
2042		.iomode = lseg->pls_range.iomode,
2043		.offset = 0,
2044		.length = NFS4_MAX_UINT64,
2045	};
2046	bool return_now = false;
2047
2048	spin_lock(&inode->i_lock);
2049	if (!pnfs_layout_is_valid(lo)) {
2050		spin_unlock(&inode->i_lock);
2051		return;
2052	}
2053	pnfs_set_plh_return_info(lo, range.iomode, 0);
2054	/*
2055	 * mark all matching lsegs so that we are sure to have no live
2056	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2057	 * for how it works.
2058	 */
2059	if (!pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0)) {
2060		nfs4_stateid stateid;
2061		enum pnfs_iomode iomode;
2062
2063		return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2064		spin_unlock(&inode->i_lock);
2065		if (return_now)
2066			pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2067	} else {
2068		spin_unlock(&inode->i_lock);
2069		nfs_commit_inode(inode, 0);
2070	}
2071}
2072EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2073
2074void
2075pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2076{
2077	if (pgio->pg_lseg == NULL ||
2078	    test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2079		return;
2080	pnfs_put_lseg(pgio->pg_lseg);
2081	pgio->pg_lseg = NULL;
2082}
2083EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2084
2085/*
2086 * Check for any intersection between the request and the pgio->pg_lseg,
2087 * and if none, put this pgio->pg_lseg away.
2088 */
2089static void
2090pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2091{
2092	if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2093		pnfs_put_lseg(pgio->pg_lseg);
2094		pgio->pg_lseg = NULL;
2095	}
2096}
2097
2098void
2099pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2100{
2101	u64 rd_size = req->wb_bytes;
2102
2103	pnfs_generic_pg_check_layout(pgio);
2104	pnfs_generic_pg_check_range(pgio, req);
2105	if (pgio->pg_lseg == NULL) {
2106		if (pgio->pg_dreq == NULL)
2107			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2108		else
2109			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2110
2111		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2112						   req->wb_context,
2113						   req_offset(req),
2114						   rd_size,
2115						   IOMODE_READ,
2116						   false,
2117						   GFP_KERNEL);
2118		if (IS_ERR(pgio->pg_lseg)) {
2119			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2120			pgio->pg_lseg = NULL;
2121			return;
2122		}
2123	}
2124	/* If no lseg, fall back to read through mds */
2125	if (pgio->pg_lseg == NULL)
2126		nfs_pageio_reset_read_mds(pgio);
2127
2128}
2129EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2130
2131void
2132pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2133			   struct nfs_page *req, u64 wb_size)
2134{
2135	pnfs_generic_pg_check_layout(pgio);
2136	pnfs_generic_pg_check_range(pgio, req);
2137	if (pgio->pg_lseg == NULL) {
2138		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2139						   req->wb_context,
2140						   req_offset(req),
2141						   wb_size,
2142						   IOMODE_RW,
2143						   false,
2144						   GFP_NOFS);
2145		if (IS_ERR(pgio->pg_lseg)) {
2146			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2147			pgio->pg_lseg = NULL;
2148			return;
2149		}
2150	}
2151	/* If no lseg, fall back to write through mds */
2152	if (pgio->pg_lseg == NULL)
2153		nfs_pageio_reset_write_mds(pgio);
2154}
2155EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2156
2157void
2158pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2159{
2160	if (desc->pg_lseg) {
2161		pnfs_put_lseg(desc->pg_lseg);
2162		desc->pg_lseg = NULL;
2163	}
2164}
2165EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2166
2167/*
2168 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2169 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2170 */
2171size_t
2172pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2173		     struct nfs_page *prev, struct nfs_page *req)
2174{
2175	unsigned int size;
2176	u64 seg_end, req_start, seg_left;
2177
2178	size = nfs_generic_pg_test(pgio, prev, req);
2179	if (!size)
2180		return 0;
2181
2182	/*
2183	 * 'size' contains the number of bytes left in the current page (up
2184	 * to the original size asked for in @req->wb_bytes).
2185	 *
2186	 * Calculate how many bytes are left in the layout segment
2187	 * and if there are less bytes than 'size', return that instead.
2188	 *
2189	 * Please also note that 'end_offset' is actually the offset of the
2190	 * first byte that lies outside the pnfs_layout_range. FIXME?
2191	 *
2192	 */
2193	if (pgio->pg_lseg) {
2194		seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2195				     pgio->pg_lseg->pls_range.length);
2196		req_start = req_offset(req);
2197
2198		/* start of request is past the last byte of this segment */
2199		if (req_start >= seg_end)
2200			return 0;
2201
2202		/* adjust 'size' iff there are fewer bytes left in the
2203		 * segment than what nfs_generic_pg_test returned */
2204		seg_left = seg_end - req_start;
2205		if (seg_left < size)
2206			size = (unsigned int)seg_left;
2207	}
2208
2209	return size;
2210}
2211EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2212
2213int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2214{
2215	struct nfs_pageio_descriptor pgio;
2216
2217	/* Resend all requests through the MDS */
2218	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2219			      hdr->completion_ops);
2220	set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2221	return nfs_pageio_resend(&pgio, hdr);
2222}
2223EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2224
2225static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2226{
2227
2228	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2229	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2230	    PNFS_LAYOUTRET_ON_ERROR) {
2231		pnfs_return_layout(hdr->inode);
2232	}
2233	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2234		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2235}
2236
2237/*
2238 * Called by non rpc-based layout drivers
2239 */
2240void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2241{
2242	if (likely(!hdr->pnfs_error)) {
2243		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2244				hdr->mds_offset + hdr->res.count);
2245		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2246	}
2247	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2248	if (unlikely(hdr->pnfs_error))
2249		pnfs_ld_handle_write_error(hdr);
2250	hdr->mds_ops->rpc_release(hdr);
2251}
2252EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2253
2254static void
2255pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2256		struct nfs_pgio_header *hdr)
2257{
2258	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2259
2260	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2261		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2262		nfs_pageio_reset_write_mds(desc);
2263		mirror->pg_recoalesce = 1;
2264	}
2265	hdr->completion_ops->completion(hdr);
2266}
2267
2268static enum pnfs_try_status
2269pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2270			const struct rpc_call_ops *call_ops,
2271			struct pnfs_layout_segment *lseg,
2272			int how)
2273{
2274	struct inode *inode = hdr->inode;
2275	enum pnfs_try_status trypnfs;
2276	struct nfs_server *nfss = NFS_SERVER(inode);
2277
2278	hdr->mds_ops = call_ops;
2279
2280	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2281		inode->i_ino, hdr->args.count, hdr->args.offset, how);
2282	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2283	if (trypnfs != PNFS_NOT_ATTEMPTED)
2284		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2285	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2286	return trypnfs;
2287}
2288
2289static void
2290pnfs_do_write(struct nfs_pageio_descriptor *desc,
2291	      struct nfs_pgio_header *hdr, int how)
2292{
2293	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2294	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2295	enum pnfs_try_status trypnfs;
2296
2297	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2298	switch (trypnfs) {
2299	case PNFS_NOT_ATTEMPTED:
2300		pnfs_write_through_mds(desc, hdr);
2301	case PNFS_ATTEMPTED:
2302		break;
2303	case PNFS_TRY_AGAIN:
2304		/* cleanup hdr and prepare to redo pnfs */
2305		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2306			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2307			list_splice_init(&hdr->pages, &mirror->pg_list);
2308			mirror->pg_recoalesce = 1;
2309		}
2310		hdr->mds_ops->rpc_release(hdr);
2311	}
2312}
2313
2314static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2315{
2316	pnfs_put_lseg(hdr->lseg);
2317	nfs_pgio_header_free(hdr);
2318}
2319
2320int
2321pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2322{
2323	struct nfs_pgio_header *hdr;
2324	int ret;
2325
2326	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2327	if (!hdr) {
2328		desc->pg_error = -ENOMEM;
2329		return desc->pg_error;
2330	}
2331	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2332
2333	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2334	ret = nfs_generic_pgio(desc, hdr);
2335	if (!ret)
2336		pnfs_do_write(desc, hdr, desc->pg_ioflags);
2337
2338	return ret;
2339}
2340EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2341
2342int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2343{
2344	struct nfs_pageio_descriptor pgio;
2345
2346	/* Resend all requests through the MDS */
2347	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2348	return nfs_pageio_resend(&pgio, hdr);
2349}
2350EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2351
2352static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2353{
2354	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2355	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2356	    PNFS_LAYOUTRET_ON_ERROR) {
2357		pnfs_return_layout(hdr->inode);
2358	}
2359	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2360		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2361}
2362
2363/*
2364 * Called by non rpc-based layout drivers
2365 */
2366void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2367{
2368	if (likely(!hdr->pnfs_error))
2369		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2370	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2371	if (unlikely(hdr->pnfs_error))
2372		pnfs_ld_handle_read_error(hdr);
2373	hdr->mds_ops->rpc_release(hdr);
2374}
2375EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2376
2377static void
2378pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2379		struct nfs_pgio_header *hdr)
2380{
2381	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2382
2383	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2384		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2385		nfs_pageio_reset_read_mds(desc);
2386		mirror->pg_recoalesce = 1;
2387	}
2388	hdr->completion_ops->completion(hdr);
2389}
2390
2391/*
2392 * Call the appropriate parallel I/O subsystem read function.
2393 */
2394static enum pnfs_try_status
2395pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2396		       const struct rpc_call_ops *call_ops,
2397		       struct pnfs_layout_segment *lseg)
2398{
2399	struct inode *inode = hdr->inode;
2400	struct nfs_server *nfss = NFS_SERVER(inode);
2401	enum pnfs_try_status trypnfs;
2402
2403	hdr->mds_ops = call_ops;
2404
2405	dprintk("%s: Reading ino:%lu %u@%llu\n",
2406		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2407
2408	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2409	if (trypnfs != PNFS_NOT_ATTEMPTED)
2410		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2411	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2412	return trypnfs;
2413}
2414
2415/* Resend all requests through pnfs. */
2416void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2417{
2418	struct nfs_pageio_descriptor pgio;
2419
2420	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2421		/* Prevent deadlocks with layoutreturn! */
2422		pnfs_put_lseg(hdr->lseg);
2423		hdr->lseg = NULL;
2424
2425		nfs_pageio_init_read(&pgio, hdr->inode, false,
2426					hdr->completion_ops);
2427		hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2428	}
2429}
2430EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2431
2432static void
2433pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2434{
2435	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2436	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2437	enum pnfs_try_status trypnfs;
2438
2439	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2440	switch (trypnfs) {
2441	case PNFS_NOT_ATTEMPTED:
2442		pnfs_read_through_mds(desc, hdr);
2443	case PNFS_ATTEMPTED:
2444		break;
2445	case PNFS_TRY_AGAIN:
2446		/* cleanup hdr and prepare to redo pnfs */
2447		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2448			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2449			list_splice_init(&hdr->pages, &mirror->pg_list);
2450			mirror->pg_recoalesce = 1;
2451		}
2452		hdr->mds_ops->rpc_release(hdr);
2453	}
2454}
2455
2456static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2457{
2458	pnfs_put_lseg(hdr->lseg);
2459	nfs_pgio_header_free(hdr);
2460}
2461
2462int
2463pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2464{
2465	struct nfs_pgio_header *hdr;
2466	int ret;
2467
2468	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2469	if (!hdr) {
2470		desc->pg_error = -ENOMEM;
2471		return desc->pg_error;
2472	}
2473	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2474	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2475	ret = nfs_generic_pgio(desc, hdr);
2476	if (!ret)
2477		pnfs_do_read(desc, hdr);
2478	return ret;
2479}
2480EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2481
2482static void pnfs_clear_layoutcommitting(struct inode *inode)
2483{
2484	unsigned long *bitlock = &NFS_I(inode)->flags;
2485
2486	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2487	smp_mb__after_atomic();
2488	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2489}
2490
2491/*
2492 * There can be multiple RW segments.
2493 */
2494static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2495{
2496	struct pnfs_layout_segment *lseg;
2497
2498	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2499		if (lseg->pls_range.iomode == IOMODE_RW &&
2500		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2501			list_add(&lseg->pls_lc_list, listp);
2502	}
2503}
2504
2505static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2506{
2507	struct pnfs_layout_segment *lseg, *tmp;
2508
2509	/* Matched by references in pnfs_set_layoutcommit */
2510	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2511		list_del_init(&lseg->pls_lc_list);
2512		pnfs_put_lseg(lseg);
2513	}
2514
2515	pnfs_clear_layoutcommitting(inode);
2516}
2517
2518void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2519{
2520	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2521}
2522EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2523
2524void
2525pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2526		loff_t end_pos)
2527{
2528	struct nfs_inode *nfsi = NFS_I(inode);
2529	bool mark_as_dirty = false;
2530
2531	spin_lock(&inode->i_lock);
2532	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2533		nfsi->layout->plh_lwb = end_pos;
2534		mark_as_dirty = true;
2535		dprintk("%s: Set layoutcommit for inode %lu ",
2536			__func__, inode->i_ino);
2537	} else if (end_pos > nfsi->layout->plh_lwb)
2538		nfsi->layout->plh_lwb = end_pos;
2539	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2540		/* references matched in nfs4_layoutcommit_release */
2541		pnfs_get_lseg(lseg);
2542	}
2543	spin_unlock(&inode->i_lock);
2544	dprintk("%s: lseg %p end_pos %llu\n",
2545		__func__, lseg, nfsi->layout->plh_lwb);
2546
2547	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2548	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2549	if (mark_as_dirty)
2550		mark_inode_dirty_sync(inode);
2551}
2552EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2553
2554void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2555{
2556	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2557
2558	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2559		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2560	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2561}
2562
2563/*
2564 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2565 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2566 * data to disk to allow the server to recover the data if it crashes.
2567 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2568 * is off, and a COMMIT is sent to a data server, or
2569 * if WRITEs to a data server return NFS_DATA_SYNC.
2570 */
2571int
2572pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2573{
2574	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2575	struct nfs4_layoutcommit_data *data;
2576	struct nfs_inode *nfsi = NFS_I(inode);
2577	loff_t end_pos;
2578	int status;
2579
2580	if (!pnfs_layoutcommit_outstanding(inode))
2581		return 0;
2582
2583	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2584
2585	status = -EAGAIN;
2586	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2587		if (!sync)
2588			goto out;
2589		status = wait_on_bit_lock_action(&nfsi->flags,
2590				NFS_INO_LAYOUTCOMMITTING,
2591				nfs_wait_bit_killable,
2592				TASK_KILLABLE);
2593		if (status)
2594			goto out;
2595	}
2596
2597	status = -ENOMEM;
2598	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2599	data = kzalloc(sizeof(*data), GFP_NOFS);
2600	if (!data)
2601		goto clear_layoutcommitting;
2602
2603	status = 0;
2604	spin_lock(&inode->i_lock);
2605	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2606		goto out_unlock;
2607
2608	INIT_LIST_HEAD(&data->lseg_list);
2609	pnfs_list_write_lseg(inode, &data->lseg_list);
2610
2611	end_pos = nfsi->layout->plh_lwb;
2612
2613	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2614	spin_unlock(&inode->i_lock);
2615
2616	data->args.inode = inode;
2617	data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2618	nfs_fattr_init(&data->fattr);
2619	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2620	data->res.fattr = &data->fattr;
2621	if (end_pos != 0)
2622		data->args.lastbytewritten = end_pos - 1;
2623	else
2624		data->args.lastbytewritten = U64_MAX;
2625	data->res.server = NFS_SERVER(inode);
2626
2627	if (ld->prepare_layoutcommit) {
2628		status = ld->prepare_layoutcommit(&data->args);
2629		if (status) {
2630			put_rpccred(data->cred);
2631			spin_lock(&inode->i_lock);
2632			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2633			if (end_pos > nfsi->layout->plh_lwb)
2634				nfsi->layout->plh_lwb = end_pos;
2635			goto out_unlock;
2636		}
2637	}
2638
2639
2640	status = nfs4_proc_layoutcommit(data, sync);
2641out:
2642	if (status)
2643		mark_inode_dirty_sync(inode);
2644	dprintk("<-- %s status %d\n", __func__, status);
2645	return status;
2646out_unlock:
2647	spin_unlock(&inode->i_lock);
2648	kfree(data);
2649clear_layoutcommitting:
2650	pnfs_clear_layoutcommitting(inode);
2651	goto out;
2652}
2653EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2654
2655int
2656pnfs_generic_sync(struct inode *inode, bool datasync)
2657{
2658	return pnfs_layoutcommit_inode(inode, true);
2659}
2660EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2661
2662struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2663{
2664	struct nfs4_threshold *thp;
2665
2666	thp = kzalloc(sizeof(*thp), GFP_NOFS);
2667	if (!thp) {
2668		dprintk("%s mdsthreshold allocation failed\n", __func__);
2669		return NULL;
2670	}
2671	return thp;
2672}
2673
2674#if IS_ENABLED(CONFIG_NFS_V4_2)
2675int
2676pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2677{
2678	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2679	struct nfs_server *server = NFS_SERVER(inode);
2680	struct nfs_inode *nfsi = NFS_I(inode);
2681	struct nfs42_layoutstat_data *data;
2682	struct pnfs_layout_hdr *hdr;
2683	int status = 0;
2684
2685	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2686		goto out;
2687
2688	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2689		goto out;
2690
2691	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2692		goto out;
2693
2694	spin_lock(&inode->i_lock);
2695	if (!NFS_I(inode)->layout) {
2696		spin_unlock(&inode->i_lock);
2697		goto out_clear_layoutstats;
2698	}
2699	hdr = NFS_I(inode)->layout;
2700	pnfs_get_layout_hdr(hdr);
2701	spin_unlock(&inode->i_lock);
2702
2703	data = kzalloc(sizeof(*data), gfp_flags);
2704	if (!data) {
2705		status = -ENOMEM;
2706		goto out_put;
2707	}
2708
2709	data->args.fh = NFS_FH(inode);
2710	data->args.inode = inode;
2711	status = ld->prepare_layoutstats(&data->args);
2712	if (status)
2713		goto out_free;
2714
2715	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2716
2717out:
2718	dprintk("%s returns %d\n", __func__, status);
2719	return status;
2720
2721out_free:
2722	kfree(data);
2723out_put:
2724	pnfs_put_layout_hdr(hdr);
2725out_clear_layoutstats:
2726	smp_mb__before_atomic();
2727	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2728	smp_mb__after_atomic();
2729	goto out;
2730}
2731EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2732#endif
2733
2734unsigned int layoutstats_timer;
2735module_param(layoutstats_timer, uint, 0644);
2736EXPORT_SYMBOL_GPL(layoutstats_timer);
v5.4
   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);
  64
  65/* Return the registered pnfs layout driver module matching given id */
  66static struct pnfs_layoutdriver_type *
  67find_pnfs_driver_locked(u32 id)
  68{
  69	struct pnfs_layoutdriver_type *local;
  70
  71	list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  72		if (local->id == id)
  73			goto out;
  74	local = NULL;
  75out:
  76	dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  77	return local;
  78}
  79
  80static struct pnfs_layoutdriver_type *
  81find_pnfs_driver(u32 id)
  82{
  83	struct pnfs_layoutdriver_type *local;
  84
  85	spin_lock(&pnfs_spinlock);
  86	local = find_pnfs_driver_locked(id);
  87	if (local != NULL && !try_module_get(local->owner)) {
  88		dprintk("%s: Could not grab reference on module\n", __func__);
  89		local = NULL;
  90	}
  91	spin_unlock(&pnfs_spinlock);
  92	return local;
  93}
  94
  95void
  96unset_pnfs_layoutdriver(struct nfs_server *nfss)
  97{
  98	if (nfss->pnfs_curr_ld) {
  99		if (nfss->pnfs_curr_ld->clear_layoutdriver)
 100			nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 101		/* Decrement the MDS count. Purge the deviceid cache if zero */
 102		if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 103			nfs4_deviceid_purge_client(nfss->nfs_client);
 104		module_put(nfss->pnfs_curr_ld->owner);
 105	}
 106	nfss->pnfs_curr_ld = NULL;
 107}
 108
 109/*
 110 * When the server sends a list of layout types, we choose one in the order
 111 * given in the list below.
 112 *
 113 * FIXME: should this list be configurable in some fashion? module param?
 114 * 	  mount option? something else?
 115 */
 116static const u32 ld_prefs[] = {
 117	LAYOUT_SCSI,
 118	LAYOUT_BLOCK_VOLUME,
 119	LAYOUT_OSD2_OBJECTS,
 120	LAYOUT_FLEX_FILES,
 121	LAYOUT_NFSV4_1_FILES,
 122	0
 123};
 124
 125static int
 126ld_cmp(const void *e1, const void *e2)
 127{
 128	u32 ld1 = *((u32 *)e1);
 129	u32 ld2 = *((u32 *)e2);
 130	int i;
 131
 132	for (i = 0; ld_prefs[i] != 0; i++) {
 133		if (ld1 == ld_prefs[i])
 134			return -1;
 135
 136		if (ld2 == ld_prefs[i])
 137			return 1;
 138	}
 139	return 0;
 140}
 141
 142/*
 143 * Try to set the server's pnfs module to the pnfs layout type specified by id.
 144 * Currently only one pNFS layout driver per filesystem is supported.
 145 *
 146 * @ids array of layout types supported by MDS.
 147 */
 148void
 149set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 150		      struct nfs_fsinfo *fsinfo)
 151{
 152	struct pnfs_layoutdriver_type *ld_type = NULL;
 153	u32 id;
 154	int i;
 155
 156	if (fsinfo->nlayouttypes == 0)
 157		goto out_no_driver;
 158	if (!(server->nfs_client->cl_exchange_flags &
 159		 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 160		printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 161			__func__, server->nfs_client->cl_exchange_flags);
 162		goto out_no_driver;
 163	}
 164
 165	sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 166		sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 167
 168	for (i = 0; i < fsinfo->nlayouttypes; i++) {
 169		id = fsinfo->layouttype[i];
 170		ld_type = find_pnfs_driver(id);
 171		if (!ld_type) {
 172			request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 173					id);
 174			ld_type = find_pnfs_driver(id);
 175		}
 176		if (ld_type)
 177			break;
 178	}
 179
 180	if (!ld_type) {
 181		dprintk("%s: No pNFS module found!\n", __func__);
 182		goto out_no_driver;
 183	}
 184
 185	server->pnfs_curr_ld = ld_type;
 186	if (ld_type->set_layoutdriver
 187	    && ld_type->set_layoutdriver(server, mntfh)) {
 188		printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 189			"driver %u.\n", __func__, id);
 190		module_put(ld_type->owner);
 191		goto out_no_driver;
 192	}
 193	/* Bump the MDS count */
 194	atomic_inc(&server->nfs_client->cl_mds_count);
 195
 196	dprintk("%s: pNFS module for %u set\n", __func__, id);
 197	return;
 198
 199out_no_driver:
 200	dprintk("%s: Using NFSv4 I/O\n", __func__);
 201	server->pnfs_curr_ld = NULL;
 202}
 203
 204int
 205pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 206{
 207	int status = -EINVAL;
 208	struct pnfs_layoutdriver_type *tmp;
 209
 210	if (ld_type->id == 0) {
 211		printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 212		return status;
 213	}
 214	if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 215		printk(KERN_ERR "NFS: %s Layout driver must provide "
 216		       "alloc_lseg and free_lseg.\n", __func__);
 217		return status;
 218	}
 219
 220	spin_lock(&pnfs_spinlock);
 221	tmp = find_pnfs_driver_locked(ld_type->id);
 222	if (!tmp) {
 223		list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 224		status = 0;
 225		dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 226			ld_type->name);
 227	} else {
 228		printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 229			__func__, ld_type->id);
 230	}
 231	spin_unlock(&pnfs_spinlock);
 232
 233	return status;
 234}
 235EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 236
 237void
 238pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 239{
 240	dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 241	spin_lock(&pnfs_spinlock);
 242	list_del(&ld_type->pnfs_tblid);
 243	spin_unlock(&pnfs_spinlock);
 244}
 245EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 246
 247/*
 248 * pNFS client layout cache
 249 */
 250
 251/* Need to hold i_lock if caller does not already hold reference */
 252void
 253pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 254{
 255	refcount_inc(&lo->plh_refcount);
 256}
 257
 258static struct pnfs_layout_hdr *
 259pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 260{
 261	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 262	return ld->alloc_layout_hdr(ino, gfp_flags);
 263}
 264
 265static void
 266pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 267{
 268	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 269	struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 270
 271	if (!list_empty(&lo->plh_layouts)) {
 272		struct nfs_client *clp = server->nfs_client;
 273
 274		spin_lock(&clp->cl_lock);
 275		list_del_init(&lo->plh_layouts);
 276		spin_unlock(&clp->cl_lock);
 277	}
 278	put_cred(lo->plh_lc_cred);
 279	return ld->free_layout_hdr(lo);
 280}
 281
 282static void
 283pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 284{
 285	struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 286	dprintk("%s: freeing layout cache %p\n", __func__, lo);
 287	nfsi->layout = NULL;
 288	/* Reset MDS Threshold I/O counters */
 289	nfsi->write_io = 0;
 290	nfsi->read_io = 0;
 291}
 292
 293void
 294pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 295{
 296	struct inode *inode;
 297
 298	if (!lo)
 299		return;
 300	inode = lo->plh_inode;
 301	pnfs_layoutreturn_before_put_layout_hdr(lo);
 302
 303	if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 304		if (!list_empty(&lo->plh_segs))
 305			WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 306		pnfs_detach_layout_hdr(lo);
 307		spin_unlock(&inode->i_lock);
 308		pnfs_free_layout_hdr(lo);
 309	}
 310}
 311
 312static void
 313pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 314			 u32 seq)
 315{
 316	if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 317		iomode = IOMODE_ANY;
 318	lo->plh_return_iomode = iomode;
 319	set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 320	if (seq != 0) {
 321		WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 322		lo->plh_return_seq = seq;
 323	}
 324}
 325
 326static void
 327pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 328{
 329	struct pnfs_layout_segment *lseg;
 330	lo->plh_return_iomode = 0;
 331	lo->plh_return_seq = 0;
 332	clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 333	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 334		if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 335			continue;
 336		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 337	}
 338}
 339
 340static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 341{
 342	clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 343	clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 344	smp_mb__after_atomic();
 345	wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 346	rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 347}
 348
 349static void
 350pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 351		struct list_head *free_me)
 352{
 353	clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 354	clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 355	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 356		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 357	if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 358		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 359}
 360
 361/*
 362 * Update the seqid of a layout stateid after receiving
 363 * NFS4ERR_OLD_STATEID
 364 */
 365bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
 366		struct pnfs_layout_range *dst_range,
 367		struct inode *inode)
 368{
 369	struct pnfs_layout_hdr *lo;
 370	struct pnfs_layout_range range = {
 371		.iomode = IOMODE_ANY,
 372		.offset = 0,
 373		.length = NFS4_MAX_UINT64,
 374	};
 375	bool ret = false;
 376	LIST_HEAD(head);
 377	int err;
 378
 379	spin_lock(&inode->i_lock);
 380	lo = NFS_I(inode)->layout;
 381	if (lo &&  pnfs_layout_is_valid(lo) &&
 382	    nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 383		/* Is our call using the most recent seqid? If so, bump it */
 384		if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
 385			nfs4_stateid_seqid_inc(dst);
 386			ret = true;
 387			goto out;
 388		}
 389		/* Try to update the seqid to the most recent */
 390		err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 391		if (err != -EBUSY) {
 392			dst->seqid = lo->plh_stateid.seqid;
 393			*dst_range = range;
 394			ret = true;
 395		}
 396	}
 397out:
 398	spin_unlock(&inode->i_lock);
 399	pnfs_free_lseg_list(&head);
 400	return ret;
 401}
 402
 403/*
 404 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 405 *
 406 * In order to continue using the pnfs_layout_hdr, a full recovery
 407 * is required.
 408 * Note that caller must hold inode->i_lock.
 409 */
 410int
 411pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 412		struct list_head *lseg_list)
 413{
 414	struct pnfs_layout_range range = {
 415		.iomode = IOMODE_ANY,
 416		.offset = 0,
 417		.length = NFS4_MAX_UINT64,
 418	};
 419	struct pnfs_layout_segment *lseg, *next;
 420
 421	set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 422	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 423		pnfs_clear_lseg_state(lseg, lseg_list);
 424	pnfs_clear_layoutreturn_info(lo);
 425	pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 426	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 427	    !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 428		pnfs_clear_layoutreturn_waitbit(lo);
 429	return !list_empty(&lo->plh_segs);
 430}
 431
 432static int
 433pnfs_iomode_to_fail_bit(u32 iomode)
 434{
 435	return iomode == IOMODE_RW ?
 436		NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 437}
 438
 439static void
 440pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 441{
 442	lo->plh_retry_timestamp = jiffies;
 443	if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 444		refcount_inc(&lo->plh_refcount);
 445}
 446
 447static void
 448pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 449{
 450	if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 451		refcount_dec(&lo->plh_refcount);
 452}
 453
 454static void
 455pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 456{
 457	struct inode *inode = lo->plh_inode;
 458	struct pnfs_layout_range range = {
 459		.iomode = iomode,
 460		.offset = 0,
 461		.length = NFS4_MAX_UINT64,
 462	};
 463	LIST_HEAD(head);
 464
 465	spin_lock(&inode->i_lock);
 466	pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 467	pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 468	spin_unlock(&inode->i_lock);
 469	pnfs_free_lseg_list(&head);
 470	dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 471			iomode == IOMODE_RW ?  "RW" : "READ");
 472}
 473
 474static bool
 475pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 476{
 477	unsigned long start, end;
 478	int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 479
 480	if (test_bit(fail_bit, &lo->plh_flags) == 0)
 481		return false;
 482	end = jiffies;
 483	start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 484	if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 485		/* It is time to retry the failed layoutgets */
 486		pnfs_layout_clear_fail_bit(lo, fail_bit);
 487		return false;
 488	}
 489	return true;
 490}
 491
 492static void
 493pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 494		const struct pnfs_layout_range *range,
 495		const nfs4_stateid *stateid)
 496{
 497	INIT_LIST_HEAD(&lseg->pls_list);
 498	INIT_LIST_HEAD(&lseg->pls_lc_list);
 499	refcount_set(&lseg->pls_refcount, 1);
 500	set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 501	lseg->pls_layout = lo;
 502	lseg->pls_range = *range;
 503	lseg->pls_seq = be32_to_cpu(stateid->seqid);
 504}
 505
 506static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 507{
 508	if (lseg != NULL) {
 509		struct inode *inode = lseg->pls_layout->plh_inode;
 510		NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 511	}
 512}
 513
 514static void
 515pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 516		struct pnfs_layout_segment *lseg)
 517{
 518	WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 519	list_del_init(&lseg->pls_list);
 520	/* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 521	refcount_dec(&lo->plh_refcount);
 522	if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 523		return;
 524	if (list_empty(&lo->plh_segs) &&
 525	    !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 526	    !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 527		if (atomic_read(&lo->plh_outstanding) == 0)
 528			set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 529		clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 530	}
 531}
 532
 533static bool
 534pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 535		struct pnfs_layout_segment *lseg)
 536{
 537	if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 538	    pnfs_layout_is_valid(lo)) {
 539		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 540		list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 541		return true;
 542	}
 543	return false;
 544}
 545
 546void
 547pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 548{
 549	struct pnfs_layout_hdr *lo;
 550	struct inode *inode;
 551
 552	if (!lseg)
 553		return;
 554
 555	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 556		refcount_read(&lseg->pls_refcount),
 557		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 558
 559	lo = lseg->pls_layout;
 560	inode = lo->plh_inode;
 561
 562	if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 563		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 564			spin_unlock(&inode->i_lock);
 565			return;
 566		}
 567		pnfs_get_layout_hdr(lo);
 568		pnfs_layout_remove_lseg(lo, lseg);
 569		if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 570			lseg = NULL;
 571		spin_unlock(&inode->i_lock);
 572		pnfs_free_lseg(lseg);
 573		pnfs_put_layout_hdr(lo);
 574	}
 575}
 576EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 577
 578/*
 579 * is l2 fully contained in l1?
 580 *   start1                             end1
 581 *   [----------------------------------)
 582 *           start2           end2
 583 *           [----------------)
 584 */
 585static bool
 586pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 587		 const struct pnfs_layout_range *l2)
 588{
 589	u64 start1 = l1->offset;
 590	u64 end1 = pnfs_end_offset(start1, l1->length);
 591	u64 start2 = l2->offset;
 592	u64 end2 = pnfs_end_offset(start2, l2->length);
 593
 594	return (start1 <= start2) && (end1 >= end2);
 595}
 596
 597static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 598		struct list_head *tmp_list)
 599{
 600	if (!refcount_dec_and_test(&lseg->pls_refcount))
 601		return false;
 602	pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 603	list_add(&lseg->pls_list, tmp_list);
 604	return true;
 605}
 606
 607/* Returns 1 if lseg is removed from list, 0 otherwise */
 608static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 609			     struct list_head *tmp_list)
 610{
 611	int rv = 0;
 612
 613	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 614		/* Remove the reference keeping the lseg in the
 615		 * list.  It will now be removed when all
 616		 * outstanding io is finished.
 617		 */
 618		dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 619			refcount_read(&lseg->pls_refcount));
 620		if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 621			rv = 1;
 622	}
 623	return rv;
 624}
 625
 626/*
 627 * Compare 2 layout stateid sequence ids, to see which is newer,
 628 * taking into account wraparound issues.
 629 */
 630static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 631{
 632	return (s32)(s1 - s2) > 0;
 633}
 634
 635static bool
 636pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 637		 const struct pnfs_layout_range *recall_range)
 638{
 639	return (recall_range->iomode == IOMODE_ANY ||
 640		lseg_range->iomode == recall_range->iomode) &&
 641	       pnfs_lseg_range_intersecting(lseg_range, recall_range);
 642}
 643
 644static bool
 645pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 646		const struct pnfs_layout_range *recall_range,
 647		u32 seq)
 648{
 649	if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 650		return false;
 651	if (recall_range == NULL)
 652		return true;
 653	return pnfs_should_free_range(&lseg->pls_range, recall_range);
 654}
 655
 656/**
 657 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 658 * @lo: layout header containing the lsegs
 659 * @tmp_list: list head where doomed lsegs should go
 660 * @recall_range: optional recall range argument to match (may be NULL)
 661 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 662 *
 663 * Walk the list of lsegs in the layout header, and tear down any that should
 664 * be destroyed. If "recall_range" is specified then the segment must match
 665 * that range. If "seq" is non-zero, then only match segments that were handed
 666 * out at or before that sequence.
 667 *
 668 * Returns number of matching invalid lsegs remaining in list after scanning
 669 * it and purging them.
 670 */
 671int
 672pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 673			    struct list_head *tmp_list,
 674			    const struct pnfs_layout_range *recall_range,
 675			    u32 seq)
 676{
 677	struct pnfs_layout_segment *lseg, *next;
 678	int remaining = 0;
 679
 680	dprintk("%s:Begin lo %p\n", __func__, lo);
 681
 682	if (list_empty(&lo->plh_segs))
 683		return 0;
 684	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 685		if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 686			dprintk("%s: freeing lseg %p iomode %d seq %u "
 687				"offset %llu length %llu\n", __func__,
 688				lseg, lseg->pls_range.iomode, lseg->pls_seq,
 689				lseg->pls_range.offset, lseg->pls_range.length);
 690			if (!mark_lseg_invalid(lseg, tmp_list))
 691				remaining++;
 692		}
 693	dprintk("%s:Return %i\n", __func__, remaining);
 694	return remaining;
 695}
 696
 697static void
 698pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 699		struct list_head *free_me,
 700		const struct pnfs_layout_range *range,
 701		u32 seq)
 702{
 703	struct pnfs_layout_segment *lseg, *next;
 704
 705	list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 706		if (pnfs_match_lseg_recall(lseg, range, seq))
 707			list_move_tail(&lseg->pls_list, free_me);
 708	}
 709}
 710
 711/* note free_me must contain lsegs from a single layout_hdr */
 712void
 713pnfs_free_lseg_list(struct list_head *free_me)
 714{
 715	struct pnfs_layout_segment *lseg, *tmp;
 716
 717	if (list_empty(free_me))
 718		return;
 719
 720	list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 721		list_del(&lseg->pls_list);
 722		pnfs_free_lseg(lseg);
 723	}
 724}
 725
 726void
 727pnfs_destroy_layout(struct nfs_inode *nfsi)
 728{
 729	struct pnfs_layout_hdr *lo;
 730	LIST_HEAD(tmp_list);
 731
 732	spin_lock(&nfsi->vfs_inode.i_lock);
 733	lo = nfsi->layout;
 734	if (lo) {
 735		pnfs_get_layout_hdr(lo);
 736		pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 737		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 738		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 739		spin_unlock(&nfsi->vfs_inode.i_lock);
 740		pnfs_free_lseg_list(&tmp_list);
 741		nfs_commit_inode(&nfsi->vfs_inode, 0);
 742		pnfs_put_layout_hdr(lo);
 743	} else
 744		spin_unlock(&nfsi->vfs_inode.i_lock);
 745}
 746EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 747
 748static bool
 749pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 750		struct list_head *layout_list)
 751{
 752	struct pnfs_layout_hdr *lo;
 753	bool ret = false;
 754
 755	spin_lock(&inode->i_lock);
 756	lo = NFS_I(inode)->layout;
 757	if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 758		pnfs_get_layout_hdr(lo);
 759		list_add(&lo->plh_bulk_destroy, layout_list);
 760		ret = true;
 761	}
 762	spin_unlock(&inode->i_lock);
 763	return ret;
 764}
 765
 766/* Caller must hold rcu_read_lock and clp->cl_lock */
 767static int
 768pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 769		struct nfs_server *server,
 770		struct list_head *layout_list)
 771	__must_hold(&clp->cl_lock)
 772	__must_hold(RCU)
 773{
 774	struct pnfs_layout_hdr *lo, *next;
 775	struct inode *inode;
 776
 777	list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 778		if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
 779		    test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
 780		    !list_empty(&lo->plh_bulk_destroy))
 781			continue;
 782		/* If the sb is being destroyed, just bail */
 783		if (!nfs_sb_active(server->super))
 784			break;
 785		inode = igrab(lo->plh_inode);
 786		if (inode != NULL) {
 787			list_del_init(&lo->plh_layouts);
 788			if (pnfs_layout_add_bulk_destroy_list(inode,
 789						layout_list))
 790				continue;
 791			rcu_read_unlock();
 792			spin_unlock(&clp->cl_lock);
 793			iput(inode);
 794		} else {
 795			rcu_read_unlock();
 796			spin_unlock(&clp->cl_lock);
 797			set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
 798		}
 799		nfs_sb_deactive(server->super);
 800		spin_lock(&clp->cl_lock);
 801		rcu_read_lock();
 802		return -EAGAIN;
 803	}
 804	return 0;
 805}
 806
 807static int
 808pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 809		bool is_bulk_recall)
 810{
 811	struct pnfs_layout_hdr *lo;
 812	struct inode *inode;
 813	LIST_HEAD(lseg_list);
 814	int ret = 0;
 815
 816	while (!list_empty(layout_list)) {
 817		lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 818				plh_bulk_destroy);
 819		dprintk("%s freeing layout for inode %lu\n", __func__,
 820			lo->plh_inode->i_ino);
 821		inode = lo->plh_inode;
 822
 823		pnfs_layoutcommit_inode(inode, false);
 824
 825		spin_lock(&inode->i_lock);
 826		list_del_init(&lo->plh_bulk_destroy);
 827		if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 828			if (is_bulk_recall)
 829				set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 830			ret = -EAGAIN;
 831		}
 832		spin_unlock(&inode->i_lock);
 833		pnfs_free_lseg_list(&lseg_list);
 834		/* Free all lsegs that are attached to commit buckets */
 835		nfs_commit_inode(inode, 0);
 836		pnfs_put_layout_hdr(lo);
 837		nfs_iput_and_deactive(inode);
 838	}
 839	return ret;
 840}
 841
 842int
 843pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 844		struct nfs_fsid *fsid,
 845		bool is_recall)
 846{
 847	struct nfs_server *server;
 848	LIST_HEAD(layout_list);
 849
 850	spin_lock(&clp->cl_lock);
 851	rcu_read_lock();
 852restart:
 853	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 854		if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 855			continue;
 856		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 857				server,
 858				&layout_list) != 0)
 859			goto restart;
 860	}
 861	rcu_read_unlock();
 862	spin_unlock(&clp->cl_lock);
 863
 864	if (list_empty(&layout_list))
 865		return 0;
 866	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 867}
 868
 869int
 870pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 871		bool is_recall)
 872{
 873	struct nfs_server *server;
 874	LIST_HEAD(layout_list);
 875
 876	spin_lock(&clp->cl_lock);
 877	rcu_read_lock();
 878restart:
 879	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 880		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 881					server,
 882					&layout_list) != 0)
 883			goto restart;
 884	}
 885	rcu_read_unlock();
 886	spin_unlock(&clp->cl_lock);
 887
 888	if (list_empty(&layout_list))
 889		return 0;
 890	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 891}
 892
 893/*
 894 * Called by the state manger to remove all layouts established under an
 895 * expired lease.
 896 */
 897void
 898pnfs_destroy_all_layouts(struct nfs_client *clp)
 899{
 900	nfs4_deviceid_mark_client_invalid(clp);
 901	nfs4_deviceid_purge_client(clp);
 902
 903	pnfs_destroy_layouts_byclid(clp, false);
 904}
 905
 906/* update lo->plh_stateid with new if is more recent */
 907void
 908pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 909			bool update_barrier)
 910{
 911	u32 oldseq, newseq, new_barrier = 0;
 912
 913	oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 914	newseq = be32_to_cpu(new->seqid);
 915
 916	if (!pnfs_layout_is_valid(lo)) {
 917		nfs4_stateid_copy(&lo->plh_stateid, new);
 918		lo->plh_barrier = newseq;
 919		pnfs_clear_layoutreturn_info(lo);
 920		clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 921		return;
 922	}
 923	if (pnfs_seqid_is_newer(newseq, oldseq)) {
 924		nfs4_stateid_copy(&lo->plh_stateid, new);
 925		/*
 926		 * Because of wraparound, we want to keep the barrier
 927		 * "close" to the current seqids.
 928		 */
 929		new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 930	}
 931	if (update_barrier)
 932		new_barrier = be32_to_cpu(new->seqid);
 933	else if (new_barrier == 0)
 934		return;
 935	if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 936		lo->plh_barrier = new_barrier;
 937}
 938
 939static bool
 940pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 941		const nfs4_stateid *stateid)
 942{
 943	u32 seqid = be32_to_cpu(stateid->seqid);
 944
 945	return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 946}
 947
 948/* lget is set to 1 if called from inside send_layoutget call chain */
 949static bool
 950pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 951{
 952	return lo->plh_block_lgets ||
 953		test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 954}
 955
 956static struct nfs_server *
 957pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
 958{
 959	struct nfs_server *server;
 960
 961	if (inode) {
 962		server = NFS_SERVER(inode);
 963	} else {
 964		struct dentry *parent_dir = dget_parent(ctx->dentry);
 965		server = NFS_SERVER(parent_dir->d_inode);
 966		dput(parent_dir);
 967	}
 968	return server;
 969}
 970
 971static void nfs4_free_pages(struct page **pages, size_t size)
 972{
 973	int i;
 974
 975	if (!pages)
 976		return;
 977
 978	for (i = 0; i < size; i++) {
 979		if (!pages[i])
 980			break;
 981		__free_page(pages[i]);
 982	}
 983	kfree(pages);
 984}
 985
 986static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
 987{
 988	struct page **pages;
 989	int i;
 990
 991	pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
 992	if (!pages) {
 993		dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
 994		return NULL;
 995	}
 996
 997	for (i = 0; i < size; i++) {
 998		pages[i] = alloc_page(gfp_flags);
 999		if (!pages[i]) {
1000			dprintk("%s: failed to allocate page\n", __func__);
1001			nfs4_free_pages(pages, i);
1002			return NULL;
1003		}
1004	}
1005
1006	return pages;
1007}
1008
1009static struct nfs4_layoutget *
1010pnfs_alloc_init_layoutget_args(struct inode *ino,
1011	   struct nfs_open_context *ctx,
1012	   const nfs4_stateid *stateid,
1013	   const struct pnfs_layout_range *range,
1014	   gfp_t gfp_flags)
1015{
1016	struct nfs_server *server = pnfs_find_server(ino, ctx);
1017	size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1018	size_t max_pages = max_response_pages(server);
1019	struct nfs4_layoutget *lgp;
 
1020
1021	dprintk("--> %s\n", __func__);
1022
 
 
 
 
 
1023	lgp = kzalloc(sizeof(*lgp), gfp_flags);
1024	if (lgp == NULL)
1025		return NULL;
1026
1027	if (max_reply_sz) {
1028		size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1029		if (npages < max_pages)
1030			max_pages = npages;
1031	}
1032
1033	lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1034	if (!lgp->args.layout.pages) {
1035		kfree(lgp);
1036		return NULL;
1037	}
1038	lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1039	lgp->res.layoutp = &lgp->args.layout;
1040
1041	/* Don't confuse uninitialised result and success */
1042	lgp->res.status = -NFS4ERR_DELAY;
1043
1044	lgp->args.minlength = PAGE_SIZE;
1045	if (lgp->args.minlength > range->length)
1046		lgp->args.minlength = range->length;
1047	if (ino) {
1048		loff_t i_size = i_size_read(ino);
1049
1050		if (range->iomode == IOMODE_READ) {
1051			if (range->offset >= i_size)
1052				lgp->args.minlength = 0;
1053			else if (i_size - range->offset < lgp->args.minlength)
1054				lgp->args.minlength = i_size - range->offset;
1055		}
1056	}
1057	lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1058	pnfs_copy_range(&lgp->args.range, range);
1059	lgp->args.type = server->pnfs_curr_ld->id;
1060	lgp->args.inode = ino;
1061	lgp->args.ctx = get_nfs_open_context(ctx);
1062	nfs4_stateid_copy(&lgp->args.stateid, stateid);
1063	lgp->gfp_flags = gfp_flags;
1064	lgp->cred = get_cred(ctx->cred);
1065	return lgp;
1066}
1067
1068void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1069{
1070	size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1071
1072	nfs4_free_pages(lgp->args.layout.pages, max_pages);
1073	if (lgp->args.inode)
1074		pnfs_put_layout_hdr(NFS_I(lgp->args.inode)->layout);
1075	put_cred(lgp->cred);
1076	put_nfs_open_context(lgp->args.ctx);
1077	kfree(lgp);
1078}
1079
1080static void pnfs_clear_layoutcommit(struct inode *inode,
1081		struct list_head *head)
1082{
1083	struct nfs_inode *nfsi = NFS_I(inode);
1084	struct pnfs_layout_segment *lseg, *tmp;
1085
1086	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1087		return;
1088	list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1089		if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1090			continue;
1091		pnfs_lseg_dec_and_remove_zero(lseg, head);
1092	}
1093}
1094
1095void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1096		const nfs4_stateid *arg_stateid,
1097		const struct pnfs_layout_range *range,
1098		const nfs4_stateid *stateid)
1099{
1100	struct inode *inode = lo->plh_inode;
1101	LIST_HEAD(freeme);
1102
1103	spin_lock(&inode->i_lock);
1104	if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
1105	    !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1106		goto out_unlock;
1107	if (stateid) {
1108		u32 seq = be32_to_cpu(arg_stateid->seqid);
1109
1110		pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1111		pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1112		pnfs_set_layout_stateid(lo, stateid, true);
1113	} else
1114		pnfs_mark_layout_stateid_invalid(lo, &freeme);
1115out_unlock:
1116	pnfs_clear_layoutreturn_waitbit(lo);
1117	spin_unlock(&inode->i_lock);
1118	pnfs_free_lseg_list(&freeme);
1119
1120}
1121
1122static bool
1123pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1124		nfs4_stateid *stateid,
1125		enum pnfs_iomode *iomode)
1126{
1127	/* Serialise LAYOUTGET/LAYOUTRETURN */
1128	if (atomic_read(&lo->plh_outstanding) != 0)
1129		return false;
1130	if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1131		return false;
1132	set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1133	pnfs_get_layout_hdr(lo);
1134	if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1135		if (stateid != NULL) {
1136			nfs4_stateid_copy(stateid, &lo->plh_stateid);
1137			if (lo->plh_return_seq != 0)
1138				stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1139		}
1140		if (iomode != NULL)
1141			*iomode = lo->plh_return_iomode;
1142		pnfs_clear_layoutreturn_info(lo);
1143		return true;
1144	}
1145	if (stateid != NULL)
1146		nfs4_stateid_copy(stateid, &lo->plh_stateid);
1147	if (iomode != NULL)
1148		*iomode = IOMODE_ANY;
1149	return true;
1150}
1151
1152static void
1153pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1154		struct pnfs_layout_hdr *lo,
1155		const nfs4_stateid *stateid,
1156		enum pnfs_iomode iomode)
1157{
1158	struct inode *inode = lo->plh_inode;
1159
1160	args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1161	args->inode = inode;
1162	args->range.iomode = iomode;
1163	args->range.offset = 0;
1164	args->range.length = NFS4_MAX_UINT64;
1165	args->layout = lo;
1166	nfs4_stateid_copy(&args->stateid, stateid);
1167}
1168
1169static int
1170pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1171		       enum pnfs_iomode iomode, bool sync)
1172{
1173	struct inode *ino = lo->plh_inode;
1174	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1175	struct nfs4_layoutreturn *lrp;
1176	int status = 0;
1177
1178	lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1179	if (unlikely(lrp == NULL)) {
1180		status = -ENOMEM;
1181		spin_lock(&ino->i_lock);
1182		pnfs_clear_layoutreturn_waitbit(lo);
1183		spin_unlock(&ino->i_lock);
1184		pnfs_put_layout_hdr(lo);
1185		goto out;
1186	}
1187
1188	pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1189	lrp->args.ld_private = &lrp->ld_private;
1190	lrp->clp = NFS_SERVER(ino)->nfs_client;
1191	lrp->cred = lo->plh_lc_cred;
1192	if (ld->prepare_layoutreturn)
1193		ld->prepare_layoutreturn(&lrp->args);
1194
1195	status = nfs4_proc_layoutreturn(lrp, sync);
1196out:
1197	dprintk("<-- %s status: %d\n", __func__, status);
1198	return status;
1199}
1200
1201/* Return true if layoutreturn is needed */
1202static bool
1203pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1204{
1205	struct pnfs_layout_segment *s;
1206	enum pnfs_iomode iomode;
1207	u32 seq;
1208
1209	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1210		return false;
1211
1212	seq = lo->plh_return_seq;
1213	iomode = lo->plh_return_iomode;
1214
1215	/* Defer layoutreturn until all recalled lsegs are done */
1216	list_for_each_entry(s, &lo->plh_segs, pls_list) {
1217		if (seq && pnfs_seqid_is_newer(s->pls_seq, seq))
1218			continue;
1219		if (iomode != IOMODE_ANY && s->pls_range.iomode != iomode)
1220			continue;
1221		if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1222			return false;
1223	}
1224
1225	return true;
1226}
1227
1228static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1229{
1230	struct inode *inode= lo->plh_inode;
1231
1232	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1233		return;
1234	spin_lock(&inode->i_lock);
1235	if (pnfs_layout_need_return(lo)) {
1236		nfs4_stateid stateid;
1237		enum pnfs_iomode iomode;
1238		bool send;
1239
1240		send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1241		spin_unlock(&inode->i_lock);
1242		if (send) {
1243			/* Send an async layoutreturn so we dont deadlock */
1244			pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1245		}
1246	} else
1247		spin_unlock(&inode->i_lock);
1248}
1249
1250/*
1251 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1252 * when the layout segment list is empty.
1253 *
1254 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1255 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1256 * deviceid is marked invalid.
1257 */
1258int
1259_pnfs_return_layout(struct inode *ino)
1260{
1261	struct pnfs_layout_hdr *lo = NULL;
1262	struct nfs_inode *nfsi = NFS_I(ino);
1263	LIST_HEAD(tmp_list);
1264	nfs4_stateid stateid;
1265	int status = 0;
1266	bool send, valid_layout;
1267
1268	dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1269
1270	spin_lock(&ino->i_lock);
1271	lo = nfsi->layout;
1272	if (!lo) {
1273		spin_unlock(&ino->i_lock);
1274		dprintk("NFS: %s no layout to return\n", __func__);
1275		goto out;
1276	}
1277	/* Reference matched in nfs4_layoutreturn_release */
1278	pnfs_get_layout_hdr(lo);
1279	/* Is there an outstanding layoutreturn ? */
1280	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1281		spin_unlock(&ino->i_lock);
1282		if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1283					TASK_UNINTERRUPTIBLE))
1284			goto out_put_layout_hdr;
1285		spin_lock(&ino->i_lock);
1286	}
1287	valid_layout = pnfs_layout_is_valid(lo);
1288	pnfs_clear_layoutcommit(ino, &tmp_list);
1289	pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1290
1291	if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1292		struct pnfs_layout_range range = {
1293			.iomode		= IOMODE_ANY,
1294			.offset		= 0,
1295			.length		= NFS4_MAX_UINT64,
1296		};
1297		NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1298	}
1299
1300	/* Don't send a LAYOUTRETURN if list was initially empty */
1301	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1302			!valid_layout) {
1303		spin_unlock(&ino->i_lock);
1304		dprintk("NFS: %s no layout segments to return\n", __func__);
1305		goto out_put_layout_hdr;
1306	}
1307
1308	send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1309	spin_unlock(&ino->i_lock);
1310	if (send)
1311		status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1312out_put_layout_hdr:
1313	pnfs_free_lseg_list(&tmp_list);
1314	pnfs_put_layout_hdr(lo);
1315out:
1316	dprintk("<-- %s status: %d\n", __func__, status);
1317	return status;
1318}
1319
1320int
1321pnfs_commit_and_return_layout(struct inode *inode)
1322{
1323	struct pnfs_layout_hdr *lo;
1324	int ret;
1325
1326	spin_lock(&inode->i_lock);
1327	lo = NFS_I(inode)->layout;
1328	if (lo == NULL) {
1329		spin_unlock(&inode->i_lock);
1330		return 0;
1331	}
1332	pnfs_get_layout_hdr(lo);
1333	/* Block new layoutgets and read/write to ds */
1334	lo->plh_block_lgets++;
1335	spin_unlock(&inode->i_lock);
1336	filemap_fdatawait(inode->i_mapping);
1337	ret = pnfs_layoutcommit_inode(inode, true);
1338	if (ret == 0)
1339		ret = _pnfs_return_layout(inode);
1340	spin_lock(&inode->i_lock);
1341	lo->plh_block_lgets--;
1342	spin_unlock(&inode->i_lock);
1343	pnfs_put_layout_hdr(lo);
1344	return ret;
1345}
1346
1347bool pnfs_roc(struct inode *ino,
1348		struct nfs4_layoutreturn_args *args,
1349		struct nfs4_layoutreturn_res *res,
1350		const struct cred *cred)
1351{
1352	struct nfs_inode *nfsi = NFS_I(ino);
1353	struct nfs_open_context *ctx;
1354	struct nfs4_state *state;
1355	struct pnfs_layout_hdr *lo;
1356	struct pnfs_layout_segment *lseg, *next;
1357	nfs4_stateid stateid;
1358	enum pnfs_iomode iomode = 0;
1359	bool layoutreturn = false, roc = false;
1360	bool skip_read = false;
1361
1362	if (!nfs_have_layout(ino))
1363		return false;
1364retry:
1365	rcu_read_lock();
1366	spin_lock(&ino->i_lock);
1367	lo = nfsi->layout;
1368	if (!lo || !pnfs_layout_is_valid(lo) ||
1369	    test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1370		lo = NULL;
1371		goto out_noroc;
1372	}
1373	pnfs_get_layout_hdr(lo);
1374	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1375		spin_unlock(&ino->i_lock);
1376		rcu_read_unlock();
1377		wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1378				TASK_UNINTERRUPTIBLE);
1379		pnfs_put_layout_hdr(lo);
1380		goto retry;
1381	}
1382
1383	/* no roc if we hold a delegation */
1384	if (nfs4_check_delegation(ino, FMODE_READ)) {
1385		if (nfs4_check_delegation(ino, FMODE_WRITE))
1386			goto out_noroc;
1387		skip_read = true;
1388	}
1389
1390	list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1391		state = ctx->state;
1392		if (state == NULL)
1393			continue;
1394		/* Don't return layout if there is open file state */
1395		if (state->state & FMODE_WRITE)
1396			goto out_noroc;
1397		if (state->state & FMODE_READ)
1398			skip_read = true;
1399	}
1400
1401
1402	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1403		if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1404			continue;
1405		/* If we are sending layoutreturn, invalidate all valid lsegs */
1406		if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1407			continue;
1408		/*
1409		 * Note: mark lseg for return so pnfs_layout_remove_lseg
1410		 * doesn't invalidate the layout for us.
1411		 */
1412		set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1413		if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1414			continue;
1415		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1416	}
1417
1418	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1419		goto out_noroc;
1420
1421	/* ROC in two conditions:
1422	 * 1. there are ROC lsegs
1423	 * 2. we don't send layoutreturn
1424	 */
1425	/* lo ref dropped in pnfs_roc_release() */
1426	layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1427	/* If the creds don't match, we can't compound the layoutreturn */
1428	if (!layoutreturn || cred != lo->plh_lc_cred)
1429		goto out_noroc;
1430
1431	roc = layoutreturn;
1432	pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1433	res->lrs_present = 0;
1434	layoutreturn = false;
1435
1436out_noroc:
1437	spin_unlock(&ino->i_lock);
1438	rcu_read_unlock();
1439	pnfs_layoutcommit_inode(ino, true);
1440	if (roc) {
1441		struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1442		if (ld->prepare_layoutreturn)
1443			ld->prepare_layoutreturn(args);
1444		pnfs_put_layout_hdr(lo);
1445		return true;
1446	}
1447	if (layoutreturn)
1448		pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1449	pnfs_put_layout_hdr(lo);
1450	return false;
1451}
1452
1453int pnfs_roc_done(struct rpc_task *task, struct inode *inode,
1454		struct nfs4_layoutreturn_args **argpp,
1455		struct nfs4_layoutreturn_res **respp,
1456		int *ret)
1457{
1458	struct nfs4_layoutreturn_args *arg = *argpp;
1459	int retval = -EAGAIN;
1460
1461	if (!arg)
1462		return 0;
1463	/* Handle Layoutreturn errors */
1464	switch (*ret) {
1465	case 0:
1466		retval = 0;
1467		break;
1468	case -NFS4ERR_NOMATCHING_LAYOUT:
1469		/* Was there an RPC level error? If not, retry */
1470		if (task->tk_rpc_status == 0)
1471			break;
1472		/* If the call was not sent, let caller handle it */
1473		if (!RPC_WAS_SENT(task))
1474			return 0;
1475		/*
1476		 * Otherwise, assume the call succeeded and
1477		 * that we need to release the layout
1478		 */
1479		*ret = 0;
1480		(*respp)->lrs_present = 0;
1481		retval = 0;
1482		break;
1483	case -NFS4ERR_DELAY:
1484		/* Let the caller handle the retry */
1485		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1486		return 0;
1487	case -NFS4ERR_OLD_STATEID:
1488		if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1489					&arg->range, inode))
1490			break;
1491		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1492		return -EAGAIN;
1493	}
1494	*argpp = NULL;
1495	*respp = NULL;
1496	return retval;
1497}
1498
1499void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1500		struct nfs4_layoutreturn_res *res,
1501		int ret)
1502{
1503	struct pnfs_layout_hdr *lo = args->layout;
1504	const nfs4_stateid *arg_stateid = NULL;
1505	const nfs4_stateid *res_stateid = NULL;
1506	struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1507
1508	switch (ret) {
1509	case -NFS4ERR_NOMATCHING_LAYOUT:
1510		break;
1511	case 0:
1512		if (res->lrs_present)
1513			res_stateid = &res->stateid;
1514		/* Fallthrough */
1515	default:
1516		arg_stateid = &args->stateid;
1517	}
1518	pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1519			res_stateid);
1520	if (ld_private && ld_private->ops && ld_private->ops->free)
1521		ld_private->ops->free(ld_private);
1522	pnfs_put_layout_hdr(lo);
1523	trace_nfs4_layoutreturn_on_close(args->inode, 0);
1524}
1525
1526bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1527{
1528	struct nfs_inode *nfsi = NFS_I(ino);
1529        struct pnfs_layout_hdr *lo;
1530        bool sleep = false;
1531
1532	/* we might not have grabbed lo reference. so need to check under
1533	 * i_lock */
1534        spin_lock(&ino->i_lock);
1535        lo = nfsi->layout;
1536        if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1537                rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1538                sleep = true;
1539	}
1540        spin_unlock(&ino->i_lock);
1541        return sleep;
1542}
1543
1544/*
1545 * Compare two layout segments for sorting into layout cache.
1546 * We want to preferentially return RW over RO layouts, so ensure those
1547 * are seen first.
1548 */
1549static s64
1550pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1551	   const struct pnfs_layout_range *l2)
1552{
1553	s64 d;
1554
1555	/* high offset > low offset */
1556	d = l1->offset - l2->offset;
1557	if (d)
1558		return d;
1559
1560	/* short length > long length */
1561	d = l2->length - l1->length;
1562	if (d)
1563		return d;
1564
1565	/* read > read/write */
1566	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1567}
1568
1569static bool
1570pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1571		const struct pnfs_layout_range *l2)
1572{
1573	return pnfs_lseg_range_cmp(l1, l2) > 0;
1574}
1575
1576static bool
1577pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1578		struct pnfs_layout_segment *old)
1579{
1580	return false;
1581}
1582
1583void
1584pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1585		   struct pnfs_layout_segment *lseg,
1586		   bool (*is_after)(const struct pnfs_layout_range *,
1587			   const struct pnfs_layout_range *),
1588		   bool (*do_merge)(struct pnfs_layout_segment *,
1589			   struct pnfs_layout_segment *),
1590		   struct list_head *free_me)
1591{
1592	struct pnfs_layout_segment *lp, *tmp;
1593
1594	dprintk("%s:Begin\n", __func__);
1595
1596	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1597		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1598			continue;
1599		if (do_merge(lseg, lp)) {
1600			mark_lseg_invalid(lp, free_me);
1601			continue;
1602		}
1603		if (is_after(&lseg->pls_range, &lp->pls_range))
1604			continue;
1605		list_add_tail(&lseg->pls_list, &lp->pls_list);
1606		dprintk("%s: inserted lseg %p "
1607			"iomode %d offset %llu length %llu before "
1608			"lp %p iomode %d offset %llu length %llu\n",
1609			__func__, lseg, lseg->pls_range.iomode,
1610			lseg->pls_range.offset, lseg->pls_range.length,
1611			lp, lp->pls_range.iomode, lp->pls_range.offset,
1612			lp->pls_range.length);
1613		goto out;
1614	}
1615	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1616	dprintk("%s: inserted lseg %p "
1617		"iomode %d offset %llu length %llu at tail\n",
1618		__func__, lseg, lseg->pls_range.iomode,
1619		lseg->pls_range.offset, lseg->pls_range.length);
1620out:
1621	pnfs_get_layout_hdr(lo);
1622
1623	dprintk("%s:Return\n", __func__);
1624}
1625EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1626
1627static void
1628pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1629		   struct pnfs_layout_segment *lseg,
1630		   struct list_head *free_me)
1631{
1632	struct inode *inode = lo->plh_inode;
1633	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1634
1635	if (ld->add_lseg != NULL)
1636		ld->add_lseg(lo, lseg, free_me);
1637	else
1638		pnfs_generic_layout_insert_lseg(lo, lseg,
1639				pnfs_lseg_range_is_after,
1640				pnfs_lseg_no_merge,
1641				free_me);
1642}
1643
1644static struct pnfs_layout_hdr *
1645alloc_init_layout_hdr(struct inode *ino,
1646		      struct nfs_open_context *ctx,
1647		      gfp_t gfp_flags)
1648{
1649	struct pnfs_layout_hdr *lo;
1650
1651	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1652	if (!lo)
1653		return NULL;
1654	refcount_set(&lo->plh_refcount, 1);
1655	INIT_LIST_HEAD(&lo->plh_layouts);
1656	INIT_LIST_HEAD(&lo->plh_segs);
1657	INIT_LIST_HEAD(&lo->plh_return_segs);
1658	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1659	lo->plh_inode = ino;
1660	lo->plh_lc_cred = get_cred(ctx->cred);
1661	lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1662	return lo;
1663}
1664
1665static struct pnfs_layout_hdr *
1666pnfs_find_alloc_layout(struct inode *ino,
1667		       struct nfs_open_context *ctx,
1668		       gfp_t gfp_flags)
1669	__releases(&ino->i_lock)
1670	__acquires(&ino->i_lock)
1671{
1672	struct nfs_inode *nfsi = NFS_I(ino);
1673	struct pnfs_layout_hdr *new = NULL;
1674
1675	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1676
1677	if (nfsi->layout != NULL)
1678		goto out_existing;
1679	spin_unlock(&ino->i_lock);
1680	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1681	spin_lock(&ino->i_lock);
1682
1683	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1684		nfsi->layout = new;
1685		return new;
1686	} else if (new != NULL)
1687		pnfs_free_layout_hdr(new);
1688out_existing:
1689	pnfs_get_layout_hdr(nfsi->layout);
1690	return nfsi->layout;
1691}
1692
1693/*
1694 * iomode matching rules:
1695 * iomode	lseg	strict match
1696 *                      iomode
1697 * -----	-----	------ -----
1698 * ANY		READ	N/A    true
1699 * ANY		RW	N/A    true
1700 * RW		READ	N/A    false
1701 * RW		RW	N/A    true
1702 * READ		READ	N/A    true
1703 * READ		RW	true   false
1704 * READ		RW	false  true
1705 */
1706static bool
1707pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1708		 const struct pnfs_layout_range *range,
1709		 bool strict_iomode)
1710{
1711	struct pnfs_layout_range range1;
1712
1713	if ((range->iomode == IOMODE_RW &&
1714	     ls_range->iomode != IOMODE_RW) ||
1715	    (range->iomode != ls_range->iomode &&
1716	     strict_iomode) ||
1717	    !pnfs_lseg_range_intersecting(ls_range, range))
1718		return false;
1719
1720	/* range1 covers only the first byte in the range */
1721	range1 = *range;
1722	range1.length = 1;
1723	return pnfs_lseg_range_contained(ls_range, &range1);
1724}
1725
1726/*
1727 * lookup range in layout
1728 */
1729static struct pnfs_layout_segment *
1730pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1731		struct pnfs_layout_range *range,
1732		bool strict_iomode)
1733{
1734	struct pnfs_layout_segment *lseg, *ret = NULL;
1735
1736	dprintk("%s:Begin\n", __func__);
1737
1738	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1739		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 
1740		    pnfs_lseg_range_match(&lseg->pls_range, range,
1741					  strict_iomode)) {
1742			ret = pnfs_get_lseg(lseg);
1743			break;
1744		}
1745	}
1746
1747	dprintk("%s:Return lseg %p ref %d\n",
1748		__func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1749	return ret;
1750}
1751
1752/*
1753 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1754 * to the MDS or over pNFS
1755 *
1756 * The nfs_inode read_io and write_io fields are cumulative counters reset
1757 * when there are no layout segments. Note that in pnfs_update_layout iomode
1758 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1759 * WRITE request.
1760 *
1761 * A return of true means use MDS I/O.
1762 *
1763 * From rfc 5661:
1764 * If a file's size is smaller than the file size threshold, data accesses
1765 * SHOULD be sent to the metadata server.  If an I/O request has a length that
1766 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1767 * server.  If both file size and I/O size are provided, the client SHOULD
1768 * reach or exceed  both thresholds before sending its read or write
1769 * requests to the data server.
1770 */
1771static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1772				     struct inode *ino, int iomode)
1773{
1774	struct nfs4_threshold *t = ctx->mdsthreshold;
1775	struct nfs_inode *nfsi = NFS_I(ino);
1776	loff_t fsize = i_size_read(ino);
1777	bool size = false, size_set = false, io = false, io_set = false, ret = false;
1778
1779	if (t == NULL)
1780		return ret;
1781
1782	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1783		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1784
1785	switch (iomode) {
1786	case IOMODE_READ:
1787		if (t->bm & THRESHOLD_RD) {
1788			dprintk("%s fsize %llu\n", __func__, fsize);
1789			size_set = true;
1790			if (fsize < t->rd_sz)
1791				size = true;
1792		}
1793		if (t->bm & THRESHOLD_RD_IO) {
1794			dprintk("%s nfsi->read_io %llu\n", __func__,
1795				nfsi->read_io);
1796			io_set = true;
1797			if (nfsi->read_io < t->rd_io_sz)
1798				io = true;
1799		}
1800		break;
1801	case IOMODE_RW:
1802		if (t->bm & THRESHOLD_WR) {
1803			dprintk("%s fsize %llu\n", __func__, fsize);
1804			size_set = true;
1805			if (fsize < t->wr_sz)
1806				size = true;
1807		}
1808		if (t->bm & THRESHOLD_WR_IO) {
1809			dprintk("%s nfsi->write_io %llu\n", __func__,
1810				nfsi->write_io);
1811			io_set = true;
1812			if (nfsi->write_io < t->wr_io_sz)
1813				io = true;
1814		}
1815		break;
1816	}
1817	if (size_set && io_set) {
1818		if (size && io)
1819			ret = true;
1820	} else if (size || io)
1821		ret = true;
1822
1823	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1824	return ret;
1825}
1826
1827static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1828{
1829	/*
1830	 * send layoutcommit as it can hold up layoutreturn due to lseg
1831	 * reference
1832	 */
1833	pnfs_layoutcommit_inode(lo->plh_inode, false);
1834	return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1835				   nfs_wait_bit_killable,
1836				   TASK_KILLABLE);
1837}
1838
1839static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1840{
1841	atomic_inc(&lo->plh_outstanding);
1842}
1843
1844static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1845{
1846	if (atomic_dec_and_test(&lo->plh_outstanding))
1847		wake_up_var(&lo->plh_outstanding);
1848}
1849
1850static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1851{
1852	unsigned long *bitlock = &lo->plh_flags;
1853
1854	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1855	smp_mb__after_atomic();
1856	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1857}
1858
1859static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1860				struct nfs_server *server)
1861{
1862	if (list_empty(&lo->plh_layouts)) {
1863		struct nfs_client *clp = server->nfs_client;
1864
1865		/* The lo must be on the clp list if there is any
1866		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1867		 */
1868		spin_lock(&clp->cl_lock);
1869		if (list_empty(&lo->plh_layouts))
1870			list_add_tail(&lo->plh_layouts, &server->layouts);
1871		spin_unlock(&clp->cl_lock);
1872	}
1873}
1874
1875/*
1876 * Layout segment is retreived from the server if not cached.
1877 * The appropriate layout segment is referenced and returned to the caller.
1878 */
1879struct pnfs_layout_segment *
1880pnfs_update_layout(struct inode *ino,
1881		   struct nfs_open_context *ctx,
1882		   loff_t pos,
1883		   u64 count,
1884		   enum pnfs_iomode iomode,
1885		   bool strict_iomode,
1886		   gfp_t gfp_flags)
1887{
1888	struct pnfs_layout_range arg = {
1889		.iomode = iomode,
1890		.offset = pos,
1891		.length = count,
1892	};
1893	unsigned pg_offset;
1894	struct nfs_server *server = NFS_SERVER(ino);
1895	struct nfs_client *clp = server->nfs_client;
1896	struct pnfs_layout_hdr *lo = NULL;
1897	struct pnfs_layout_segment *lseg = NULL;
1898	struct nfs4_layoutget *lgp;
1899	nfs4_stateid stateid;
1900	long timeout = 0;
1901	unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1902	bool first;
1903
1904	if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1905		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1906				 PNFS_UPDATE_LAYOUT_NO_PNFS);
1907		goto out;
1908	}
1909
 
 
 
 
 
 
1910	if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1911		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1912				 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1913		goto out;
1914	}
1915
1916lookup_again:
1917	lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1918	if (IS_ERR(lseg))
1919		goto out;
1920	first = false;
1921	spin_lock(&ino->i_lock);
1922	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1923	if (lo == NULL) {
1924		spin_unlock(&ino->i_lock);
1925		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1926				 PNFS_UPDATE_LAYOUT_NOMEM);
1927		goto out;
1928	}
1929
1930	/* Do we even need to bother with this? */
1931	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1932		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1933				 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1934		dprintk("%s matches recall, use MDS\n", __func__);
1935		goto out_unlock;
1936	}
1937
1938	/* if LAYOUTGET already failed once we don't try again */
1939	if (pnfs_layout_io_test_failed(lo, iomode)) {
1940		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1941				 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1942		goto out_unlock;
1943	}
1944
1945	/*
1946	 * If the layout segment list is empty, but there are outstanding
1947	 * layoutget calls, then they might be subject to a layoutrecall.
1948	 */
1949	if (list_empty(&lo->plh_segs) &&
1950	    atomic_read(&lo->plh_outstanding) != 0) {
1951		spin_unlock(&ino->i_lock);
1952		lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
1953					!atomic_read(&lo->plh_outstanding)));
1954		if (IS_ERR(lseg))
1955			goto out_put_layout_hdr;
1956		pnfs_put_layout_hdr(lo);
1957		goto lookup_again;
1958	}
1959
1960	lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1961	if (lseg) {
1962		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1963				PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1964		goto out_unlock;
1965	}
1966
 
 
 
 
 
 
1967	/*
1968	 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1969	 * stateid, or it has been invalidated, then we must use the open
1970	 * stateid.
1971	 */
1972	if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1973		int status;
1974
1975		/*
1976		 * The first layoutget for the file. Need to serialize per
1977		 * RFC 5661 Errata 3208.
1978		 */
1979		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1980				     &lo->plh_flags)) {
1981			spin_unlock(&ino->i_lock);
1982			lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
1983						NFS_LAYOUT_FIRST_LAYOUTGET,
1984						TASK_KILLABLE));
1985			if (IS_ERR(lseg))
1986				goto out_put_layout_hdr;
1987			pnfs_put_layout_hdr(lo);
1988			dprintk("%s retrying\n", __func__);
1989			goto lookup_again;
1990		}
1991
1992		first = true;
1993		status = nfs4_select_rw_stateid(ctx->state,
1994					iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
1995					NULL, &stateid, NULL);
1996		if (status != 0) {
1997			lseg = ERR_PTR(status);
1998			trace_pnfs_update_layout(ino, pos, count,
1999					iomode, lo, lseg,
2000					PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2001			if (status != -EAGAIN)
2002				goto out_unlock;
2003			spin_unlock(&ino->i_lock);
2004			nfs4_schedule_stateid_recovery(server, ctx->state);
2005			pnfs_clear_first_layoutget(lo);
2006			pnfs_put_layout_hdr(lo);
2007			goto lookup_again;
2008		}
2009	} else {
2010		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2011	}
2012
2013	/*
2014	 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
2015	 * for LAYOUTRETURN even if first is true.
2016	 */
2017	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2018		spin_unlock(&ino->i_lock);
2019		dprintk("%s wait for layoutreturn\n", __func__);
2020		lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2021		if (!IS_ERR(lseg)) {
2022			if (first)
2023				pnfs_clear_first_layoutget(lo);
2024			pnfs_put_layout_hdr(lo);
2025			dprintk("%s retrying\n", __func__);
2026			trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2027					lseg, PNFS_UPDATE_LAYOUT_RETRY);
2028			goto lookup_again;
2029		}
2030		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2031				PNFS_UPDATE_LAYOUT_RETURN);
2032		goto out_put_layout_hdr;
2033	}
2034
2035	if (pnfs_layoutgets_blocked(lo)) {
2036		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2037				PNFS_UPDATE_LAYOUT_BLOCKED);
2038		goto out_unlock;
2039	}
2040	nfs_layoutget_begin(lo);
2041	spin_unlock(&ino->i_lock);
2042
2043	_add_to_server_list(lo, server);
 
 
 
 
 
 
 
 
2044
2045	pg_offset = arg.offset & ~PAGE_MASK;
2046	if (pg_offset) {
2047		arg.offset -= pg_offset;
2048		arg.length += pg_offset;
2049	}
2050	if (arg.length != NFS4_MAX_UINT64)
2051		arg.length = PAGE_ALIGN(arg.length);
2052
2053	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2054	if (!lgp) {
2055		trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2056					 PNFS_UPDATE_LAYOUT_NOMEM);
2057		nfs_layoutget_end(lo);
2058		goto out_put_layout_hdr;
2059	}
2060
2061	lseg = nfs4_proc_layoutget(lgp, &timeout);
2062	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2063				 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2064	nfs_layoutget_end(lo);
2065	if (IS_ERR(lseg)) {
2066		switch(PTR_ERR(lseg)) {
2067		case -EBUSY:
2068			if (time_after(jiffies, giveup))
2069				lseg = NULL;
2070			break;
2071		case -ERECALLCONFLICT:
 
 
 
 
 
 
 
 
 
2072		case -EAGAIN:
2073			break;
2074		default:
2075			if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2076				pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2077				lseg = NULL;
2078			}
2079			goto out_put_layout_hdr;
2080		}
2081		if (lseg) {
2082			if (first)
2083				pnfs_clear_first_layoutget(lo);
2084			trace_pnfs_update_layout(ino, pos, count,
2085				iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2086			pnfs_put_layout_hdr(lo);
2087			goto lookup_again;
2088		}
2089	} else {
2090		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2091	}
2092
2093out_put_layout_hdr:
2094	if (first)
2095		pnfs_clear_first_layoutget(lo);
2096	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2097				 PNFS_UPDATE_LAYOUT_EXIT);
2098	pnfs_put_layout_hdr(lo);
2099out:
2100	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2101			"(%s, offset: %llu, length: %llu)\n",
2102			__func__, ino->i_sb->s_id,
2103			(unsigned long long)NFS_FILEID(ino),
2104			IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2105			iomode==IOMODE_RW ?  "read/write" : "read-only",
2106			(unsigned long long)pos,
2107			(unsigned long long)count);
2108	return lseg;
2109out_unlock:
2110	spin_unlock(&ino->i_lock);
2111	goto out_put_layout_hdr;
2112}
2113EXPORT_SYMBOL_GPL(pnfs_update_layout);
2114
2115static bool
2116pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2117{
2118	switch (range->iomode) {
2119	case IOMODE_READ:
2120	case IOMODE_RW:
2121		break;
2122	default:
2123		return false;
2124	}
2125	if (range->offset == NFS4_MAX_UINT64)
2126		return false;
2127	if (range->length == 0)
2128		return false;
2129	if (range->length != NFS4_MAX_UINT64 &&
2130	    range->length > NFS4_MAX_UINT64 - range->offset)
2131		return false;
2132	return true;
2133}
2134
2135static struct pnfs_layout_hdr *
2136_pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2137{
2138	struct pnfs_layout_hdr *lo;
2139
2140	spin_lock(&ino->i_lock);
2141	lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2142	if (!lo)
2143		goto out_unlock;
2144	if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2145		goto out_unlock;
2146	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2147		goto out_unlock;
2148	if (pnfs_layoutgets_blocked(lo))
2149		goto out_unlock;
2150	if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2151		goto out_unlock;
2152	nfs_layoutget_begin(lo);
2153	spin_unlock(&ino->i_lock);
2154	_add_to_server_list(lo, NFS_SERVER(ino));
2155	return lo;
2156
2157out_unlock:
2158	spin_unlock(&ino->i_lock);
2159	pnfs_put_layout_hdr(lo);
2160	return NULL;
2161}
2162
2163extern const nfs4_stateid current_stateid;
2164
2165static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2166				     struct nfs_open_context *ctx)
2167{
2168	struct inode *ino = data->dentry->d_inode;
2169	struct pnfs_layout_range rng = {
2170		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2171			  IOMODE_RW: IOMODE_READ,
2172		.offset = 0,
2173		.length = NFS4_MAX_UINT64,
2174	};
2175	struct nfs4_layoutget *lgp;
2176	struct pnfs_layout_hdr *lo;
2177
2178	/* Heuristic: don't send layoutget if we have cached data */
2179	if (rng.iomode == IOMODE_READ &&
2180	   (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2181		return;
2182
2183	lo = _pnfs_grab_empty_layout(ino, ctx);
2184	if (!lo)
2185		return;
2186	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2187					     &rng, GFP_KERNEL);
2188	if (!lgp) {
2189		pnfs_clear_first_layoutget(lo);
2190		pnfs_put_layout_hdr(lo);
2191		return;
2192	}
2193	data->lgp = lgp;
2194	data->o_arg.lg_args = &lgp->args;
2195	data->o_res.lg_res = &lgp->res;
2196}
2197
2198static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2199				     struct nfs_open_context *ctx)
2200{
2201	struct pnfs_layout_range rng = {
2202		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2203			  IOMODE_RW: IOMODE_READ,
2204		.offset = 0,
2205		.length = NFS4_MAX_UINT64,
2206	};
2207	struct nfs4_layoutget *lgp;
2208
2209	lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2210					     &rng, GFP_KERNEL);
2211	if (!lgp)
2212		return;
2213	data->lgp = lgp;
2214	data->o_arg.lg_args = &lgp->args;
2215	data->o_res.lg_res = &lgp->res;
2216}
2217
2218void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2219			 struct nfs_open_context *ctx)
2220{
2221	struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2222
2223	if (!(pnfs_enabled_sb(server) &&
2224	      server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2225		return;
2226	/* Could check on max_ops, but currently hardcoded high enough */
2227	if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2228		return;
2229	if (data->state)
2230		_lgopen_prepare_attached(data, ctx);
2231	else
2232		_lgopen_prepare_floating(data, ctx);
2233}
2234
2235void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2236		       struct nfs_open_context *ctx)
2237{
2238	struct pnfs_layout_hdr *lo;
2239	struct pnfs_layout_segment *lseg;
2240	struct nfs_server *srv = NFS_SERVER(ino);
2241	u32 iomode;
2242
2243	if (!lgp)
2244		return;
2245	dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2246	if (lgp->res.status) {
2247		switch (lgp->res.status) {
2248		default:
2249			break;
2250		/*
2251		 * Halt lgopen attempts if the server doesn't recognise
2252		 * the "current stateid" value, the layout type, or the
2253		 * layoutget operation as being valid.
2254		 * Also if it complains about too many ops in the compound
2255		 * or of the request/reply being too big.
2256		 */
2257		case -NFS4ERR_BAD_STATEID:
2258		case -NFS4ERR_NOTSUPP:
2259		case -NFS4ERR_REP_TOO_BIG:
2260		case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2261		case -NFS4ERR_REQ_TOO_BIG:
2262		case -NFS4ERR_TOO_MANY_OPS:
2263		case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2264			srv->caps &= ~NFS_CAP_LGOPEN;
2265		}
2266		return;
2267	}
2268	if (!lgp->args.inode) {
2269		lo = _pnfs_grab_empty_layout(ino, ctx);
2270		if (!lo)
2271			return;
2272		lgp->args.inode = ino;
2273	} else
2274		lo = NFS_I(lgp->args.inode)->layout;
2275
2276	lseg = pnfs_layout_process(lgp);
2277	if (!IS_ERR(lseg)) {
2278		iomode = lgp->args.range.iomode;
2279		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2280		pnfs_put_lseg(lseg);
2281	}
2282}
2283
2284void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2285{
2286	if (lgp != NULL) {
2287		struct inode *inode = lgp->args.inode;
2288		if (inode) {
2289			struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2290			pnfs_clear_first_layoutget(lo);
2291			nfs_layoutget_end(lo);
2292		}
2293		pnfs_layoutget_free(lgp);
2294	}
2295}
2296
2297struct pnfs_layout_segment *
2298pnfs_layout_process(struct nfs4_layoutget *lgp)
2299{
2300	struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2301	struct nfs4_layoutget_res *res = &lgp->res;
2302	struct pnfs_layout_segment *lseg;
2303	struct inode *ino = lo->plh_inode;
2304	LIST_HEAD(free_me);
2305
2306	if (!pnfs_sanity_check_layout_range(&res->range))
2307		return ERR_PTR(-EINVAL);
2308
2309	/* Inject layout blob into I/O device driver */
2310	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2311	if (IS_ERR_OR_NULL(lseg)) {
2312		if (!lseg)
2313			lseg = ERR_PTR(-ENOMEM);
2314
2315		dprintk("%s: Could not allocate layout: error %ld\n",
2316		       __func__, PTR_ERR(lseg));
2317		return lseg;
2318	}
2319
2320	pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2321
2322	spin_lock(&ino->i_lock);
2323	if (pnfs_layoutgets_blocked(lo)) {
2324		dprintk("%s forget reply due to state\n", __func__);
2325		goto out_forget;
2326	}
2327
2328	if (!pnfs_layout_is_valid(lo)) {
2329		/* We have a completely new layout */
2330		pnfs_set_layout_stateid(lo, &res->stateid, true);
2331	} else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2332		/* existing state ID, make sure the sequence number matches. */
2333		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2334			dprintk("%s forget reply due to sequence\n", __func__);
2335			goto out_forget;
2336		}
2337		pnfs_set_layout_stateid(lo, &res->stateid, false);
2338	} else {
2339		/*
2340		 * We got an entirely new state ID.  Mark all segments for the
2341		 * inode invalid, and retry the layoutget
2342		 */
2343		pnfs_mark_layout_stateid_invalid(lo, &free_me);
2344		goto out_forget;
2345	}
2346
2347	pnfs_get_lseg(lseg);
2348	pnfs_layout_insert_lseg(lo, lseg, &free_me);
2349
2350
2351	if (res->return_on_close)
2352		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2353
2354	spin_unlock(&ino->i_lock);
2355	pnfs_free_lseg_list(&free_me);
2356	return lseg;
2357
2358out_forget:
2359	spin_unlock(&ino->i_lock);
2360	lseg->pls_layout = lo;
2361	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 
 
2362	return ERR_PTR(-EAGAIN);
2363}
2364
2365static int
2366mark_lseg_invalid_or_return(struct pnfs_layout_segment *lseg,
2367		struct list_head *tmp_list)
2368{
2369	if (!mark_lseg_invalid(lseg, tmp_list))
2370		return 0;
2371	pnfs_cache_lseg_for_layoutreturn(lseg->pls_layout, lseg);
2372	return 1;
2373}
2374
2375/**
2376 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2377 * @lo: pointer to layout header
2378 * @tmp_list: list header to be used with pnfs_free_lseg_list()
2379 * @return_range: describe layout segment ranges to be returned
2380 * @seq: stateid seqid to match
2381 *
2382 * This function is mainly intended for use by layoutrecall. It attempts
2383 * to free the layout segment immediately, or else to mark it for return
2384 * as soon as its reference count drops to zero.
2385 *
2386 * Returns
2387 * - 0: a layoutreturn needs to be scheduled.
2388 * - EBUSY: there are layout segment that are still in use.
2389 * - ENOENT: there are no layout segments that need to be returned.
2390 */
2391int
2392pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2393				struct list_head *tmp_list,
2394				const struct pnfs_layout_range *return_range,
2395				u32 seq)
2396{
2397	struct pnfs_layout_segment *lseg, *next;
2398	int remaining = 0;
2399
2400	dprintk("%s:Begin lo %p\n", __func__, lo);
2401
 
 
 
2402	assert_spin_locked(&lo->plh_inode->i_lock);
2403
2404	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2405		if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2406			dprintk("%s: marking lseg %p iomode %d "
2407				"offset %llu length %llu\n", __func__,
2408				lseg, lseg->pls_range.iomode,
2409				lseg->pls_range.offset,
2410				lseg->pls_range.length);
2411			if (mark_lseg_invalid_or_return(lseg, tmp_list))
2412				continue;
2413			remaining++;
2414			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2415		}
2416
2417	if (remaining) {
2418		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2419		return -EBUSY;
2420	}
2421
2422	if (!list_empty(&lo->plh_return_segs)) {
2423		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2424		return 0;
2425	}
2426
2427	return -ENOENT;
2428}
2429
2430void pnfs_error_mark_layout_for_return(struct inode *inode,
2431				       struct pnfs_layout_segment *lseg)
2432{
2433	struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2434	struct pnfs_layout_range range = {
2435		.iomode = lseg->pls_range.iomode,
2436		.offset = 0,
2437		.length = NFS4_MAX_UINT64,
2438	};
2439	bool return_now = false;
2440
2441	spin_lock(&inode->i_lock);
2442	if (!pnfs_layout_is_valid(lo)) {
2443		spin_unlock(&inode->i_lock);
2444		return;
2445	}
2446	pnfs_set_plh_return_info(lo, range.iomode, 0);
2447	/*
2448	 * mark all matching lsegs so that we are sure to have no live
2449	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2450	 * for how it works.
2451	 */
2452	if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0) != -EBUSY) {
2453		nfs4_stateid stateid;
2454		enum pnfs_iomode iomode;
2455
2456		return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2457		spin_unlock(&inode->i_lock);
2458		if (return_now)
2459			pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2460	} else {
2461		spin_unlock(&inode->i_lock);
2462		nfs_commit_inode(inode, 0);
2463	}
2464}
2465EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2466
2467void
2468pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2469{
2470	if (pgio->pg_lseg == NULL ||
2471	    test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2472		return;
2473	pnfs_put_lseg(pgio->pg_lseg);
2474	pgio->pg_lseg = NULL;
2475}
2476EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2477
2478/*
2479 * Check for any intersection between the request and the pgio->pg_lseg,
2480 * and if none, put this pgio->pg_lseg away.
2481 */
2482static void
2483pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2484{
2485	if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2486		pnfs_put_lseg(pgio->pg_lseg);
2487		pgio->pg_lseg = NULL;
2488	}
2489}
2490
2491void
2492pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2493{
2494	u64 rd_size = req->wb_bytes;
2495
2496	pnfs_generic_pg_check_layout(pgio);
2497	pnfs_generic_pg_check_range(pgio, req);
2498	if (pgio->pg_lseg == NULL) {
2499		if (pgio->pg_dreq == NULL)
2500			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2501		else
2502			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2503
2504		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2505						   nfs_req_openctx(req),
2506						   req_offset(req),
2507						   rd_size,
2508						   IOMODE_READ,
2509						   false,
2510						   GFP_KERNEL);
2511		if (IS_ERR(pgio->pg_lseg)) {
2512			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2513			pgio->pg_lseg = NULL;
2514			return;
2515		}
2516	}
2517	/* If no lseg, fall back to read through mds */
2518	if (pgio->pg_lseg == NULL)
2519		nfs_pageio_reset_read_mds(pgio);
2520
2521}
2522EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2523
2524void
2525pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2526			   struct nfs_page *req, u64 wb_size)
2527{
2528	pnfs_generic_pg_check_layout(pgio);
2529	pnfs_generic_pg_check_range(pgio, req);
2530	if (pgio->pg_lseg == NULL) {
2531		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2532						   nfs_req_openctx(req),
2533						   req_offset(req),
2534						   wb_size,
2535						   IOMODE_RW,
2536						   false,
2537						   GFP_KERNEL);
2538		if (IS_ERR(pgio->pg_lseg)) {
2539			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2540			pgio->pg_lseg = NULL;
2541			return;
2542		}
2543	}
2544	/* If no lseg, fall back to write through mds */
2545	if (pgio->pg_lseg == NULL)
2546		nfs_pageio_reset_write_mds(pgio);
2547}
2548EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2549
2550void
2551pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2552{
2553	if (desc->pg_lseg) {
2554		pnfs_put_lseg(desc->pg_lseg);
2555		desc->pg_lseg = NULL;
2556	}
2557}
2558EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2559
2560/*
2561 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2562 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2563 */
2564size_t
2565pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2566		     struct nfs_page *prev, struct nfs_page *req)
2567{
2568	unsigned int size;
2569	u64 seg_end, req_start, seg_left;
2570
2571	size = nfs_generic_pg_test(pgio, prev, req);
2572	if (!size)
2573		return 0;
2574
2575	/*
2576	 * 'size' contains the number of bytes left in the current page (up
2577	 * to the original size asked for in @req->wb_bytes).
2578	 *
2579	 * Calculate how many bytes are left in the layout segment
2580	 * and if there are less bytes than 'size', return that instead.
2581	 *
2582	 * Please also note that 'end_offset' is actually the offset of the
2583	 * first byte that lies outside the pnfs_layout_range. FIXME?
2584	 *
2585	 */
2586	if (pgio->pg_lseg) {
2587		seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2588				     pgio->pg_lseg->pls_range.length);
2589		req_start = req_offset(req);
2590
2591		/* start of request is past the last byte of this segment */
2592		if (req_start >= seg_end)
2593			return 0;
2594
2595		/* adjust 'size' iff there are fewer bytes left in the
2596		 * segment than what nfs_generic_pg_test returned */
2597		seg_left = seg_end - req_start;
2598		if (seg_left < size)
2599			size = (unsigned int)seg_left;
2600	}
2601
2602	return size;
2603}
2604EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2605
2606int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2607{
2608	struct nfs_pageio_descriptor pgio;
2609
2610	/* Resend all requests through the MDS */
2611	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2612			      hdr->completion_ops);
2613	set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2614	return nfs_pageio_resend(&pgio, hdr);
2615}
2616EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2617
2618static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2619{
2620
2621	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2622	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2623	    PNFS_LAYOUTRET_ON_ERROR) {
2624		pnfs_return_layout(hdr->inode);
2625	}
2626	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2627		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2628}
2629
2630/*
2631 * Called by non rpc-based layout drivers
2632 */
2633void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2634{
2635	if (likely(!hdr->pnfs_error)) {
2636		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2637				hdr->mds_offset + hdr->res.count);
2638		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2639	}
2640	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2641	if (unlikely(hdr->pnfs_error))
2642		pnfs_ld_handle_write_error(hdr);
2643	hdr->mds_ops->rpc_release(hdr);
2644}
2645EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2646
2647static void
2648pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2649		struct nfs_pgio_header *hdr)
2650{
2651	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2652
2653	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2654		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2655		nfs_pageio_reset_write_mds(desc);
2656		mirror->pg_recoalesce = 1;
2657	}
2658	hdr->completion_ops->completion(hdr);
2659}
2660
2661static enum pnfs_try_status
2662pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2663			const struct rpc_call_ops *call_ops,
2664			struct pnfs_layout_segment *lseg,
2665			int how)
2666{
2667	struct inode *inode = hdr->inode;
2668	enum pnfs_try_status trypnfs;
2669	struct nfs_server *nfss = NFS_SERVER(inode);
2670
2671	hdr->mds_ops = call_ops;
2672
2673	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2674		inode->i_ino, hdr->args.count, hdr->args.offset, how);
2675	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2676	if (trypnfs != PNFS_NOT_ATTEMPTED)
2677		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2678	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2679	return trypnfs;
2680}
2681
2682static void
2683pnfs_do_write(struct nfs_pageio_descriptor *desc,
2684	      struct nfs_pgio_header *hdr, int how)
2685{
2686	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2687	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2688	enum pnfs_try_status trypnfs;
2689
2690	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2691	switch (trypnfs) {
2692	case PNFS_NOT_ATTEMPTED:
2693		pnfs_write_through_mds(desc, hdr);
2694	case PNFS_ATTEMPTED:
2695		break;
2696	case PNFS_TRY_AGAIN:
2697		/* cleanup hdr and prepare to redo pnfs */
2698		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2699			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2700			list_splice_init(&hdr->pages, &mirror->pg_list);
2701			mirror->pg_recoalesce = 1;
2702		}
2703		hdr->mds_ops->rpc_release(hdr);
2704	}
2705}
2706
2707static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2708{
2709	pnfs_put_lseg(hdr->lseg);
2710	nfs_pgio_header_free(hdr);
2711}
2712
2713int
2714pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2715{
2716	struct nfs_pgio_header *hdr;
2717	int ret;
2718
2719	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2720	if (!hdr) {
2721		desc->pg_error = -ENOMEM;
2722		return desc->pg_error;
2723	}
2724	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2725
2726	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2727	ret = nfs_generic_pgio(desc, hdr);
2728	if (!ret)
2729		pnfs_do_write(desc, hdr, desc->pg_ioflags);
2730
2731	return ret;
2732}
2733EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2734
2735int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2736{
2737	struct nfs_pageio_descriptor pgio;
2738
2739	/* Resend all requests through the MDS */
2740	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2741	return nfs_pageio_resend(&pgio, hdr);
2742}
2743EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2744
2745static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2746{
2747	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2748	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2749	    PNFS_LAYOUTRET_ON_ERROR) {
2750		pnfs_return_layout(hdr->inode);
2751	}
2752	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2753		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2754}
2755
2756/*
2757 * Called by non rpc-based layout drivers
2758 */
2759void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2760{
2761	if (likely(!hdr->pnfs_error))
2762		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2763	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2764	if (unlikely(hdr->pnfs_error))
2765		pnfs_ld_handle_read_error(hdr);
2766	hdr->mds_ops->rpc_release(hdr);
2767}
2768EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2769
2770static void
2771pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2772		struct nfs_pgio_header *hdr)
2773{
2774	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2775
2776	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2777		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2778		nfs_pageio_reset_read_mds(desc);
2779		mirror->pg_recoalesce = 1;
2780	}
2781	hdr->completion_ops->completion(hdr);
2782}
2783
2784/*
2785 * Call the appropriate parallel I/O subsystem read function.
2786 */
2787static enum pnfs_try_status
2788pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2789		       const struct rpc_call_ops *call_ops,
2790		       struct pnfs_layout_segment *lseg)
2791{
2792	struct inode *inode = hdr->inode;
2793	struct nfs_server *nfss = NFS_SERVER(inode);
2794	enum pnfs_try_status trypnfs;
2795
2796	hdr->mds_ops = call_ops;
2797
2798	dprintk("%s: Reading ino:%lu %u@%llu\n",
2799		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2800
2801	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2802	if (trypnfs != PNFS_NOT_ATTEMPTED)
2803		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2804	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2805	return trypnfs;
2806}
2807
2808/* Resend all requests through pnfs. */
2809void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2810{
2811	struct nfs_pageio_descriptor pgio;
2812
2813	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2814		/* Prevent deadlocks with layoutreturn! */
2815		pnfs_put_lseg(hdr->lseg);
2816		hdr->lseg = NULL;
2817
2818		nfs_pageio_init_read(&pgio, hdr->inode, false,
2819					hdr->completion_ops);
2820		hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2821	}
2822}
2823EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2824
2825static void
2826pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2827{
2828	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2829	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2830	enum pnfs_try_status trypnfs;
2831
2832	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2833	switch (trypnfs) {
2834	case PNFS_NOT_ATTEMPTED:
2835		pnfs_read_through_mds(desc, hdr);
2836	case PNFS_ATTEMPTED:
2837		break;
2838	case PNFS_TRY_AGAIN:
2839		/* cleanup hdr and prepare to redo pnfs */
2840		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2841			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2842			list_splice_init(&hdr->pages, &mirror->pg_list);
2843			mirror->pg_recoalesce = 1;
2844		}
2845		hdr->mds_ops->rpc_release(hdr);
2846	}
2847}
2848
2849static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2850{
2851	pnfs_put_lseg(hdr->lseg);
2852	nfs_pgio_header_free(hdr);
2853}
2854
2855int
2856pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2857{
2858	struct nfs_pgio_header *hdr;
2859	int ret;
2860
2861	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2862	if (!hdr) {
2863		desc->pg_error = -ENOMEM;
2864		return desc->pg_error;
2865	}
2866	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2867	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2868	ret = nfs_generic_pgio(desc, hdr);
2869	if (!ret)
2870		pnfs_do_read(desc, hdr);
2871	return ret;
2872}
2873EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2874
2875static void pnfs_clear_layoutcommitting(struct inode *inode)
2876{
2877	unsigned long *bitlock = &NFS_I(inode)->flags;
2878
2879	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2880	smp_mb__after_atomic();
2881	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2882}
2883
2884/*
2885 * There can be multiple RW segments.
2886 */
2887static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2888{
2889	struct pnfs_layout_segment *lseg;
2890
2891	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2892		if (lseg->pls_range.iomode == IOMODE_RW &&
2893		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2894			list_add(&lseg->pls_lc_list, listp);
2895	}
2896}
2897
2898static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2899{
2900	struct pnfs_layout_segment *lseg, *tmp;
2901
2902	/* Matched by references in pnfs_set_layoutcommit */
2903	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2904		list_del_init(&lseg->pls_lc_list);
2905		pnfs_put_lseg(lseg);
2906	}
2907
2908	pnfs_clear_layoutcommitting(inode);
2909}
2910
2911void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2912{
2913	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2914}
2915EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2916
2917void
2918pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2919		loff_t end_pos)
2920{
2921	struct nfs_inode *nfsi = NFS_I(inode);
2922	bool mark_as_dirty = false;
2923
2924	spin_lock(&inode->i_lock);
2925	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2926		nfsi->layout->plh_lwb = end_pos;
2927		mark_as_dirty = true;
2928		dprintk("%s: Set layoutcommit for inode %lu ",
2929			__func__, inode->i_ino);
2930	} else if (end_pos > nfsi->layout->plh_lwb)
2931		nfsi->layout->plh_lwb = end_pos;
2932	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2933		/* references matched in nfs4_layoutcommit_release */
2934		pnfs_get_lseg(lseg);
2935	}
2936	spin_unlock(&inode->i_lock);
2937	dprintk("%s: lseg %p end_pos %llu\n",
2938		__func__, lseg, nfsi->layout->plh_lwb);
2939
2940	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2941	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2942	if (mark_as_dirty)
2943		mark_inode_dirty_sync(inode);
2944}
2945EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2946
2947void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2948{
2949	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2950
2951	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2952		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2953	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2954}
2955
2956/*
2957 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2958 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2959 * data to disk to allow the server to recover the data if it crashes.
2960 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2961 * is off, and a COMMIT is sent to a data server, or
2962 * if WRITEs to a data server return NFS_DATA_SYNC.
2963 */
2964int
2965pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2966{
2967	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2968	struct nfs4_layoutcommit_data *data;
2969	struct nfs_inode *nfsi = NFS_I(inode);
2970	loff_t end_pos;
2971	int status;
2972
2973	if (!pnfs_layoutcommit_outstanding(inode))
2974		return 0;
2975
2976	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2977
2978	status = -EAGAIN;
2979	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2980		if (!sync)
2981			goto out;
2982		status = wait_on_bit_lock_action(&nfsi->flags,
2983				NFS_INO_LAYOUTCOMMITTING,
2984				nfs_wait_bit_killable,
2985				TASK_KILLABLE);
2986		if (status)
2987			goto out;
2988	}
2989
2990	status = -ENOMEM;
2991	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2992	data = kzalloc(sizeof(*data), GFP_NOFS);
2993	if (!data)
2994		goto clear_layoutcommitting;
2995
2996	status = 0;
2997	spin_lock(&inode->i_lock);
2998	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2999		goto out_unlock;
3000
3001	INIT_LIST_HEAD(&data->lseg_list);
3002	pnfs_list_write_lseg(inode, &data->lseg_list);
3003
3004	end_pos = nfsi->layout->plh_lwb;
3005
3006	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3007	spin_unlock(&inode->i_lock);
3008
3009	data->args.inode = inode;
3010	data->cred = get_cred(nfsi->layout->plh_lc_cred);
3011	nfs_fattr_init(&data->fattr);
3012	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3013	data->res.fattr = &data->fattr;
3014	if (end_pos != 0)
3015		data->args.lastbytewritten = end_pos - 1;
3016	else
3017		data->args.lastbytewritten = U64_MAX;
3018	data->res.server = NFS_SERVER(inode);
3019
3020	if (ld->prepare_layoutcommit) {
3021		status = ld->prepare_layoutcommit(&data->args);
3022		if (status) {
3023			put_cred(data->cred);
3024			spin_lock(&inode->i_lock);
3025			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3026			if (end_pos > nfsi->layout->plh_lwb)
3027				nfsi->layout->plh_lwb = end_pos;
3028			goto out_unlock;
3029		}
3030	}
3031
3032
3033	status = nfs4_proc_layoutcommit(data, sync);
3034out:
3035	if (status)
3036		mark_inode_dirty_sync(inode);
3037	dprintk("<-- %s status %d\n", __func__, status);
3038	return status;
3039out_unlock:
3040	spin_unlock(&inode->i_lock);
3041	kfree(data);
3042clear_layoutcommitting:
3043	pnfs_clear_layoutcommitting(inode);
3044	goto out;
3045}
3046EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3047
3048int
3049pnfs_generic_sync(struct inode *inode, bool datasync)
3050{
3051	return pnfs_layoutcommit_inode(inode, true);
3052}
3053EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3054
3055struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3056{
3057	struct nfs4_threshold *thp;
3058
3059	thp = kzalloc(sizeof(*thp), GFP_NOFS);
3060	if (!thp) {
3061		dprintk("%s mdsthreshold allocation failed\n", __func__);
3062		return NULL;
3063	}
3064	return thp;
3065}
3066
3067#if IS_ENABLED(CONFIG_NFS_V4_2)
3068int
3069pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3070{
3071	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3072	struct nfs_server *server = NFS_SERVER(inode);
3073	struct nfs_inode *nfsi = NFS_I(inode);
3074	struct nfs42_layoutstat_data *data;
3075	struct pnfs_layout_hdr *hdr;
3076	int status = 0;
3077
3078	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3079		goto out;
3080
3081	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3082		goto out;
3083
3084	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3085		goto out;
3086
3087	spin_lock(&inode->i_lock);
3088	if (!NFS_I(inode)->layout) {
3089		spin_unlock(&inode->i_lock);
3090		goto out_clear_layoutstats;
3091	}
3092	hdr = NFS_I(inode)->layout;
3093	pnfs_get_layout_hdr(hdr);
3094	spin_unlock(&inode->i_lock);
3095
3096	data = kzalloc(sizeof(*data), gfp_flags);
3097	if (!data) {
3098		status = -ENOMEM;
3099		goto out_put;
3100	}
3101
3102	data->args.fh = NFS_FH(inode);
3103	data->args.inode = inode;
3104	status = ld->prepare_layoutstats(&data->args);
3105	if (status)
3106		goto out_free;
3107
3108	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3109
3110out:
3111	dprintk("%s returns %d\n", __func__, status);
3112	return status;
3113
3114out_free:
3115	kfree(data);
3116out_put:
3117	pnfs_put_layout_hdr(hdr);
3118out_clear_layoutstats:
3119	smp_mb__before_atomic();
3120	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3121	smp_mb__after_atomic();
3122	goto out;
3123}
3124EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3125#endif
3126
3127unsigned int layoutstats_timer;
3128module_param(layoutstats_timer, uint, 0644);
3129EXPORT_SYMBOL_GPL(layoutstats_timer);