write.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * linux/fs/nfs/write.c
   4 *
   5 * Write file data over NFS.
   6 *
   7 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   8 */
   9
  10#include <linux/types.h>
  11#include <linux/slab.h>
  12#include <linux/mm.h>
  13#include <linux/pagemap.h>
  14#include <linux/file.h>
  15#include <linux/writeback.h>
  16#include <linux/swap.h>
  17#include <linux/migrate.h>
  18
  19#include <linux/sunrpc/clnt.h>
  20#include <linux/nfs_fs.h>
  21#include <linux/nfs_mount.h>
  22#include <linux/nfs_page.h>
  23#include <linux/backing-dev.h>
  24#include <linux/export.h>
  25#include <linux/freezer.h>
  26#include <linux/wait.h>
  27#include <linux/iversion.h>
  28#include <linux/filelock.h>
  29
  30#include <linux/uaccess.h>
  31#include <linux/sched/mm.h>
  32
  33#include "delegation.h"
  34#include "internal.h"
  35#include "iostat.h"
  36#include "nfs4_fs.h"
  37#include "fscache.h"
  38#include "pnfs.h"
  39
  40#include "nfstrace.h"
  41
  42#define NFSDBG_FACILITY		NFSDBG_PAGECACHE
  43
  44#define MIN_POOL_WRITE		(32)
  45#define MIN_POOL_COMMIT		(4)
  46
  47struct nfs_io_completion {
  48	void (*complete)(void *data);
  49	void *data;
  50	struct kref refcount;
  51};
  52
  53/*
  54 * Local function declarations
  55 */
  56static void nfs_redirty_request(struct nfs_page *req);
  57static const struct rpc_call_ops nfs_commit_ops;
  58static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  59static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  60static const struct nfs_rw_ops nfs_rw_write_ops;
  61static void nfs_inode_remove_request(struct nfs_page *req);
  62static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
  63				     struct nfs_page *req);
  64static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  65				      struct inode *inode);
  66
  67static struct kmem_cache *nfs_wdata_cachep;
  68static mempool_t *nfs_wdata_mempool;
  69static struct kmem_cache *nfs_cdata_cachep;
  70static mempool_t *nfs_commit_mempool;
  71
  72struct nfs_commit_data *nfs_commitdata_alloc(void)
  73{
  74	struct nfs_commit_data *p;
  75
  76	p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
  77	if (!p) {
  78		p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  79		if (!p)
  80			return NULL;
  81		memset(p, 0, sizeof(*p));
  82	}
  83	INIT_LIST_HEAD(&p->pages);
  84	return p;
  85}
  86EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  87
  88void nfs_commit_free(struct nfs_commit_data *p)
  89{
  90	mempool_free(p, nfs_commit_mempool);
  91}
  92EXPORT_SYMBOL_GPL(nfs_commit_free);
  93
  94static struct nfs_pgio_header *nfs_writehdr_alloc(void)
  95{
  96	struct nfs_pgio_header *p;
  97
  98	p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
  99	if (!p) {
 100		p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
 101		if (!p)
 102			return NULL;
 103		memset(p, 0, sizeof(*p));
 104	}
 105	p->rw_mode = FMODE_WRITE;
 106	return p;
 107}
 108
 109static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 110{
 111	mempool_free(hdr, nfs_wdata_mempool);
 112}
 113
 114static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 115{
 116	return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 117}
 118
 119static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 120		void (*complete)(void *), void *data)
 121{
 122	ioc->complete = complete;
 123	ioc->data = data;
 124	kref_init(&ioc->refcount);
 125}
 126
 127static void nfs_io_completion_release(struct kref *kref)
 128{
 129	struct nfs_io_completion *ioc = container_of(kref,
 130			struct nfs_io_completion, refcount);
 131	ioc->complete(ioc->data);
 132	kfree(ioc);
 133}
 134
 135static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 136{
 137	if (ioc != NULL)
 138		kref_get(&ioc->refcount);
 139}
 140
 141static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 142{
 143	if (ioc != NULL)
 144		kref_put(&ioc->refcount, nfs_io_completion_release);
 145}
 146
 147static void
 148nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
 149{
 150	if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
 151		kref_get(&req->wb_kref);
 152		atomic_long_inc(&NFS_I(inode)->nrequests);
 153	}
 154}
 155
 156static void nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
 157{
 158	if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
 159		nfs_page_set_inode_ref(req, inode);
 160}
 161
 162/**
 163 * nfs_folio_find_head_request - find head request associated with a folio
 164 * @folio: pointer to folio
 165 *
 166 * must be called while holding the inode lock.
 167 *
 168 * returns matching head request with reference held, or NULL if not found.
 169 */
 170static struct nfs_page *nfs_folio_find_head_request(struct folio *folio)
 171{
 172	struct address_space *mapping = folio->mapping;
 173	struct nfs_page *req;
 174
 175	if (!folio_test_private(folio))
 176		return NULL;
 177	spin_lock(&mapping->i_private_lock);
 178	req = folio->private;
 179	if (req) {
 180		WARN_ON_ONCE(req->wb_head != req);
 181		kref_get(&req->wb_kref);
 182	}
 183	spin_unlock(&mapping->i_private_lock);
 184	return req;
 185}
 186
 187/* Adjust the file length if we're writing beyond the end */
 188static void nfs_grow_file(struct folio *folio, unsigned int offset,
 189			  unsigned int count)
 190{
 191	struct inode *inode = folio->mapping->host;
 192	loff_t end, i_size;
 193	pgoff_t end_index;
 194
 195	spin_lock(&inode->i_lock);
 196	i_size = i_size_read(inode);
 197	end_index = ((i_size - 1) >> folio_shift(folio)) << folio_order(folio);
 198	if (i_size > 0 && folio->index < end_index)
 199		goto out;
 200	end = folio_pos(folio) + (loff_t)offset + (loff_t)count;
 201	if (i_size >= end)
 202		goto out;
 203	trace_nfs_size_grow(inode, end);
 204	i_size_write(inode, end);
 205	NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 206	nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 207out:
 208	/* Atomically update timestamps if they are delegated to us. */
 209	nfs_update_delegated_mtime_locked(inode);
 210	spin_unlock(&inode->i_lock);
 211	nfs_fscache_invalidate(inode, 0);
 212}
 213
 214/* A writeback failed: mark the page as bad, and invalidate the page cache */
 215static void nfs_set_pageerror(struct address_space *mapping)
 216{
 217	struct inode *inode = mapping->host;
 218
 219	nfs_zap_mapping(mapping->host, mapping);
 220	/* Force file size revalidation */
 221	spin_lock(&inode->i_lock);
 222	nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
 223					     NFS_INO_INVALID_CHANGE |
 224					     NFS_INO_INVALID_SIZE);
 225	spin_unlock(&inode->i_lock);
 226}
 227
 228static void nfs_mapping_set_error(struct folio *folio, int error)
 229{
 230	struct address_space *mapping = folio->mapping;
 231
 232	filemap_set_wb_err(mapping, error);
 233	if (mapping->host)
 234		errseq_set(&mapping->host->i_sb->s_wb_err,
 235			   error == -ENOSPC ? -ENOSPC : -EIO);
 236	nfs_set_pageerror(mapping);
 237}
 238
 239/*
 240 * nfs_page_covers_folio
 241 * @req: struct nfs_page
 242 *
 243 * Return true if the request covers the whole folio.
 244 * Note that the caller should ensure all subrequests have been joined
 245 */
 246static bool nfs_page_group_covers_page(struct nfs_page *req)
 247{
 248	unsigned int len = nfs_folio_length(nfs_page_to_folio(req));
 249
 250	return req->wb_pgbase == 0 && req->wb_bytes == len;
 251}
 252
 253/* We can set the PG_uptodate flag if we see that a write request
 254 * covers the full page.
 255 */
 256static void nfs_mark_uptodate(struct nfs_page *req)
 257{
 258	struct folio *folio = nfs_page_to_folio(req);
 259
 260	if (folio_test_uptodate(folio))
 261		return;
 262	if (!nfs_page_group_covers_page(req))
 263		return;
 264	folio_mark_uptodate(folio);
 265}
 266
 267static int wb_priority(struct writeback_control *wbc)
 268{
 269	int ret = 0;
 270
 271	if (wbc->sync_mode == WB_SYNC_ALL)
 272		ret = FLUSH_COND_STABLE;
 273	return ret;
 274}
 275
 276/*
 277 * NFS congestion control
 278 */
 279
 280int nfs_congestion_kb;
 281
 282#define NFS_CONGESTION_ON_THRESH 	(nfs_congestion_kb >> (PAGE_SHIFT-10))
 283#define NFS_CONGESTION_OFF_THRESH	\
 284	(NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 285
 286static void nfs_folio_set_writeback(struct folio *folio)
 287{
 288	struct nfs_server *nfss = NFS_SERVER(folio->mapping->host);
 289
 290	folio_start_writeback(folio);
 291	if (atomic_long_inc_return(&nfss->writeback) > NFS_CONGESTION_ON_THRESH)
 292		nfss->write_congested = 1;
 293}
 294
 295static void nfs_folio_end_writeback(struct folio *folio)
 296{
 297	struct nfs_server *nfss = NFS_SERVER(folio->mapping->host);
 298
 299	folio_end_writeback_no_dropbehind(folio);
 300	if (atomic_long_dec_return(&nfss->writeback) <
 301	    NFS_CONGESTION_OFF_THRESH) {
 302		nfss->write_congested = 0;
 303		wake_up_all(&nfss->write_congestion_wait);
 304	}
 305}
 306
 307static void nfs_page_end_writeback(struct nfs_page *req)
 308{
 309	if (nfs_page_group_sync_on_bit(req, PG_WB_END)) {
 310		nfs_unlock_request(req);
 311		nfs_folio_end_writeback(nfs_page_to_folio(req));
 312	} else
 313		nfs_unlock_request(req);
 314}
 315
 316/*
 317 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 318 *
 319 * @destroy_list - request list (using wb_this_page) terminated by @old_head
 320 * @old_head - the old head of the list
 321 *
 322 * All subrequests must be locked and removed from all lists, so at this point
 323 * they are only "active" in this function, and possibly in nfs_wait_on_request
 324 * with a reference held by some other context.
 325 */
 326static void
 327nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 328				 struct nfs_page *old_head,
 329				 struct inode *inode)
 330{
 331	while (destroy_list) {
 332		struct nfs_page *subreq = destroy_list;
 333
 334		destroy_list = (subreq->wb_this_page == old_head) ?
 335				   NULL : subreq->wb_this_page;
 336
 337		/* Note: lock subreq in order to change subreq->wb_head */
 338		nfs_page_set_headlock(subreq);
 339		WARN_ON_ONCE(old_head != subreq->wb_head);
 340
 341		/* make sure old group is not used */
 342		subreq->wb_this_page = subreq;
 343		subreq->wb_head = subreq;
 344
 345		clear_bit(PG_REMOVE, &subreq->wb_flags);
 346
 347		/* Note: races with nfs_page_group_destroy() */
 348		if (!kref_read(&subreq->wb_kref)) {
 349			/* Check if we raced with nfs_page_group_destroy() */
 350			if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 351				nfs_page_clear_headlock(subreq);
 352				nfs_free_request(subreq);
 353			} else
 354				nfs_page_clear_headlock(subreq);
 355			continue;
 356		}
 357		nfs_page_clear_headlock(subreq);
 358
 359		nfs_release_request(old_head);
 360
 361		if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 362			nfs_release_request(subreq);
 363			atomic_long_dec(&NFS_I(inode)->nrequests);
 364		}
 365
 366		/* subreq is now totally disconnected from page group or any
 367		 * write / commit lists. last chance to wake any waiters */
 368		nfs_unlock_and_release_request(subreq);
 369	}
 370}
 371
 372/*
 373 * nfs_join_page_group - destroy subrequests of the head req
 374 * @head: the page used to lookup the "page group" of nfs_page structures
 375 * @inode: Inode to which the request belongs.
 376 *
 377 * This function joins all sub requests to the head request by first
 378 * locking all requests in the group, cancelling any pending operations
 379 * and finally updating the head request to cover the whole range covered by
 380 * the (former) group.  All subrequests are removed from any write or commit
 381 * lists, unlinked from the group and destroyed.
 382 */
 383void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo,
 384			 struct inode *inode)
 385{
 386	struct nfs_page *subreq;
 387	struct nfs_page *destroy_list = NULL;
 388	unsigned int pgbase, off, bytes;
 389
 390	pgbase = head->wb_pgbase;
 391	bytes = head->wb_bytes;
 392	off = head->wb_offset;
 393	for (subreq = head->wb_this_page; subreq != head;
 394			subreq = subreq->wb_this_page) {
 395		/* Subrequests should always form a contiguous range */
 396		if (pgbase > subreq->wb_pgbase) {
 397			off -= pgbase - subreq->wb_pgbase;
 398			bytes += pgbase - subreq->wb_pgbase;
 399			pgbase = subreq->wb_pgbase;
 400		}
 401		bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 402				- pgbase, bytes);
 403	}
 404
 405	/* Set the head request's range to cover the former page group */
 406	head->wb_pgbase = pgbase;
 407	head->wb_bytes = bytes;
 408	head->wb_offset = off;
 409
 410	/* Now that all requests are locked, make sure they aren't on any list.
 411	 * Commit list removal accounting is done after locks are dropped */
 412	subreq = head;
 413	do {
 414		nfs_clear_request_commit(cinfo, subreq);
 415		subreq = subreq->wb_this_page;
 416	} while (subreq != head);
 417
 418	/* unlink subrequests from head, destroy them later */
 419	if (head->wb_this_page != head) {
 420		/* destroy list will be terminated by head */
 421		destroy_list = head->wb_this_page;
 422		head->wb_this_page = head;
 423	}
 424
 425	nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 426}
 427
 428/**
 429 * nfs_wait_on_request - Wait for a request to complete.
 430 * @req: request to wait upon.
 431 *
 432 * Interruptible by fatal signals only.
 433 * The user is responsible for holding a count on the request.
 434 */
 435static int nfs_wait_on_request(struct nfs_page *req)
 436{
 437	if (!test_bit(PG_BUSY, &req->wb_flags))
 438		return 0;
 439	set_bit(PG_CONTENDED2, &req->wb_flags);
 440	smp_mb__after_atomic();
 441	return wait_on_bit_io(&req->wb_flags, PG_BUSY,
 442			      TASK_UNINTERRUPTIBLE);
 443}
 444
 445/*
 446 * nfs_unroll_locks -  unlock all newly locked reqs and wait on @req
 447 * @head: head request of page group, must be holding head lock
 448 * @req: request that couldn't lock and needs to wait on the req bit lock
 449 *
 450 * This is a helper function for nfs_lock_and_join_requests
 451 * returns 0 on success, < 0 on error.
 452 */
 453static void
 454nfs_unroll_locks(struct nfs_page *head, struct nfs_page *req)
 455{
 456	struct nfs_page *tmp;
 457
 458	/* relinquish all the locks successfully grabbed this run */
 459	for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 460		if (!kref_read(&tmp->wb_kref))
 461			continue;
 462		nfs_unlock_and_release_request(tmp);
 463	}
 464}
 465
 466/*
 467 * nfs_page_group_lock_subreq -  try to lock a subrequest
 468 * @head: head request of page group
 469 * @subreq: request to lock
 470 *
 471 * This is a helper function for nfs_lock_and_join_requests which
 472 * must be called with the head request and page group both locked.
 473 * On error, it returns with the page group unlocked.
 474 */
 475static int
 476nfs_page_group_lock_subreq(struct nfs_page *head, struct nfs_page *subreq)
 477{
 478	int ret;
 479
 480	if (!kref_get_unless_zero(&subreq->wb_kref))
 481		return 0;
 482	while (!nfs_lock_request(subreq)) {
 483		nfs_page_group_unlock(head);
 484		ret = nfs_wait_on_request(subreq);
 485		if (!ret)
 486			ret = nfs_page_group_lock(head);
 487		if (ret < 0) {
 488			nfs_unroll_locks(head, subreq);
 489			nfs_release_request(subreq);
 490			return ret;
 491		}
 492	}
 493	return 0;
 494}
 495
 496/*
 497 * nfs_lock_and_join_requests - join all subreqs to the head req
 498 * @folio: the folio used to lookup the "page group" of nfs_page structures
 499 *
 500 * This function joins all sub requests to the head request by first
 501 * locking all requests in the group, cancelling any pending operations
 502 * and finally updating the head request to cover the whole range covered by
 503 * the (former) group.  All subrequests are removed from any write or commit
 504 * lists, unlinked from the group and destroyed.
 505 *
 506 * Returns a locked, referenced pointer to the head request - which after
 507 * this call is guaranteed to be the only request associated with the page.
 508 * Returns NULL if no requests are found for @folio, or a ERR_PTR if an
 509 * error was encountered.
 510 */
 511static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio)
 512{
 513	struct inode *inode = folio->mapping->host;
 514	struct nfs_page *head, *subreq;
 515	struct nfs_commit_info cinfo;
 516	int ret;
 517
 518	/*
 519	 * A reference is taken only on the head request which acts as a
 520	 * reference to the whole page group - the group will not be destroyed
 521	 * until the head reference is released.
 522	 */
 523retry:
 524	head = nfs_folio_find_head_request(folio);
 525	if (!head)
 526		return NULL;
 527
 528	while (!nfs_lock_request(head)) {
 529		ret = nfs_wait_on_request(head);
 530		if (ret < 0) {
 531			nfs_release_request(head);
 532			return ERR_PTR(ret);
 533		}
 534	}
 535
 536	ret = nfs_page_group_lock(head);
 537	if (ret < 0)
 538		goto out_unlock;
 539
 540	/* Ensure that nobody removed the request before we locked it */
 541	if (head != folio->private) {
 542		nfs_page_group_unlock(head);
 543		nfs_unlock_and_release_request(head);
 544		goto retry;
 545	}
 546
 547	nfs_cancel_remove_inode(head, inode);
 548
 549	/* lock each request in the page group */
 550	for (subreq = head->wb_this_page;
 551	     subreq != head;
 552	     subreq = subreq->wb_this_page) {
 553		ret = nfs_page_group_lock_subreq(head, subreq);
 554		if (ret < 0)
 555			goto out_unlock;
 556	}
 557
 558	nfs_page_group_unlock(head);
 559
 560	nfs_init_cinfo_from_inode(&cinfo, inode);
 561	nfs_join_page_group(head, &cinfo, inode);
 562	return head;
 563
 564out_unlock:
 565	nfs_unlock_and_release_request(head);
 566	return ERR_PTR(ret);
 567}
 568
 569static void nfs_write_error(struct nfs_page *req, int error)
 570{
 571	trace_nfs_write_error(nfs_page_to_inode(req), req, error);
 572	nfs_mapping_set_error(nfs_page_to_folio(req), error);
 573	nfs_inode_remove_request(req);
 574	nfs_page_end_writeback(req);
 575	nfs_release_request(req);
 576}
 577
 578/*
 579 * Find an associated nfs write request, and prepare to flush it out
 580 * May return an error if the user signalled nfs_wait_on_request().
 581 */
 582static int nfs_do_writepage(struct folio *folio, struct writeback_control *wbc,
 583		struct nfs_pageio_descriptor *pgio)
 584{
 585	struct nfs_page *req;
 586	int ret;
 587
 588	nfs_pageio_cond_complete(pgio, folio->index);
 589
 590	req = nfs_lock_and_join_requests(folio);
 591	if (!req)
 592		return 0;
 593	if (IS_ERR(req))
 594		return PTR_ERR(req);
 595
 596	trace_nfs_do_writepage(req);
 597	nfs_folio_set_writeback(folio);
 598	WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 599
 600	/* If there is a fatal error that covers this write, just exit */
 601	ret = pgio->pg_error;
 602	if (nfs_error_is_fatal_on_server(ret))
 603		goto out_launder;
 604
 605	if (!nfs_pageio_add_request(pgio, req)) {
 606		ret = pgio->pg_error;
 607		/*
 608		 * Remove the problematic req upon fatal errors on the server
 609		 */
 610		if (nfs_error_is_fatal_on_server(ret))
 611			goto out_launder;
 612		folio_redirty_for_writepage(wbc, folio);
 613		nfs_redirty_request(req);
 614		pgio->pg_error = 0;
 615		return ret;
 616	}
 617
 618	nfs_add_stats(folio->mapping->host, NFSIOS_WRITEPAGES, 1);
 619	return 0;
 620
 621out_launder:
 622	nfs_write_error(req, ret);
 623	return 0;
 624}
 625
 626/*
 627 * Write an mmapped page to the server.
 628 */
 629static int nfs_writepage_locked(struct folio *folio,
 630				struct writeback_control *wbc)
 631{
 632	struct nfs_pageio_descriptor pgio;
 633	struct inode *inode = folio->mapping->host;
 634	int err;
 635
 636	nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 637	nfs_pageio_init_write(&pgio, inode, 0, false,
 638			      &nfs_async_write_completion_ops);
 639	err = nfs_do_writepage(folio, wbc, &pgio);
 640	pgio.pg_error = 0;
 641	nfs_pageio_complete(&pgio);
 642	return err;
 643}
 644
 645static void nfs_io_completion_commit(void *inode)
 646{
 647	nfs_commit_inode(inode, 0);
 648}
 649
 650int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 651{
 652	struct inode *inode = mapping->host;
 653	struct nfs_pageio_descriptor pgio;
 654	struct nfs_io_completion *ioc = NULL;
 655	unsigned int mntflags = NFS_SERVER(inode)->flags;
 656	struct nfs_server *nfss = NFS_SERVER(inode);
 657	int priority = 0;
 658	int err;
 659
 660	trace_nfs_writepages(inode, wbc->range_start, wbc->range_end - wbc->range_start);
 661
 662	/* Wait with writeback until write congestion eases */
 663	if (wbc->sync_mode == WB_SYNC_NONE && nfss->write_congested) {
 664		err = wait_event_killable(nfss->write_congestion_wait,
 665					  nfss->write_congested == 0);
 666		if (err)
 667			goto out_err;
 668	}
 669
 670	nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 671
 672	if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 673	    wbc->for_background || wbc->for_sync) {
 674		ioc = nfs_io_completion_alloc(GFP_KERNEL);
 675		if (ioc)
 676			nfs_io_completion_init(ioc, nfs_io_completion_commit,
 677					       inode);
 678		priority = wb_priority(wbc);
 679	}
 680
 681	do {
 682		struct folio *folio = NULL;
 683
 684		nfs_pageio_init_write(&pgio, inode, priority, false,
 685				      &nfs_async_write_completion_ops);
 686		pgio.pg_io_completion = ioc;
 687		while ((folio = writeback_iter(mapping, wbc, folio, &err))) {
 688			err = nfs_do_writepage(folio, wbc, &pgio);
 689			folio_unlock(folio);
 690		}
 691		pgio.pg_error = 0;
 692		nfs_pageio_complete(&pgio);
 693		if (err == -EAGAIN && mntflags & NFS_MOUNT_SOFTERR)
 694			break;
 695	} while (err < 0 && !nfs_error_is_fatal(err));
 696	nfs_io_completion_put(ioc);
 697
 698	if (err > 0)
 699		err = 0;
 700out_err:
 701	trace_nfs_writepages_done(inode, wbc->range_start, wbc->range_end - wbc->range_start, err);
 702	return err;
 703}
 704
 705/*
 706 * Insert a write request into an inode
 707 */
 708static void nfs_inode_add_request(struct nfs_page *req)
 709{
 710	struct folio *folio = nfs_page_to_folio(req);
 711	struct address_space *mapping = folio->mapping;
 712	struct nfs_inode *nfsi = NFS_I(mapping->host);
 713
 714	WARN_ON_ONCE(req->wb_this_page != req);
 715
 716	/* Lock the request! */
 717	nfs_lock_request(req);
 718	spin_lock(&mapping->i_private_lock);
 719	set_bit(PG_MAPPED, &req->wb_flags);
 720	folio_set_private(folio);
 721	folio->private = req;
 722	spin_unlock(&mapping->i_private_lock);
 723	atomic_long_inc(&nfsi->nrequests);
 724	/* this a head request for a page group - mark it as having an
 725	 * extra reference so sub groups can follow suit.
 726	 * This flag also informs pgio layer when to bump nrequests when
 727	 * adding subrequests. */
 728	WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 729	kref_get(&req->wb_kref);
 730}
 731
 732/*
 733 * Remove a write request from an inode
 734 */
 735static void nfs_inode_remove_request(struct nfs_page *req)
 736{
 737	struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
 738
 739	nfs_page_group_lock(req);
 740	if (nfs_page_group_sync_on_bit_locked(req, PG_REMOVE)) {
 741		struct folio *folio = nfs_page_to_folio(req->wb_head);
 742		struct address_space *mapping = folio->mapping;
 743
 744		spin_lock(&mapping->i_private_lock);
 745		if (likely(folio)) {
 746			folio->private = NULL;
 747			folio_clear_private(folio);
 748			clear_bit(PG_MAPPED, &req->wb_head->wb_flags);
 749		}
 750		spin_unlock(&mapping->i_private_lock);
 751
 752		folio_end_dropbehind(folio);
 753	}
 754	nfs_page_group_unlock(req);
 755
 756	if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 757		atomic_long_dec(&nfsi->nrequests);
 758		nfs_release_request(req);
 759	}
 760}
 761
 762static void nfs_mark_request_dirty(struct nfs_page *req)
 763{
 764	struct folio *folio = nfs_page_to_folio(req);
 765	if (folio)
 766		filemap_dirty_folio(folio_mapping(folio), folio);
 767}
 768
 769/**
 770 * nfs_request_add_commit_list_locked - add request to a commit list
 771 * @req: pointer to a struct nfs_page
 772 * @dst: commit list head
 773 * @cinfo: holds list lock and accounting info
 774 *
 775 * This sets the PG_CLEAN bit, updates the cinfo count of
 776 * number of outstanding requests requiring a commit as well as
 777 * the MM page stats.
 778 *
 779 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 780 * nfs_page lock.
 781 */
 782void
 783nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 784			    struct nfs_commit_info *cinfo)
 785{
 786	set_bit(PG_CLEAN, &req->wb_flags);
 787	nfs_list_add_request(req, dst);
 788	atomic_long_inc(&cinfo->mds->ncommit);
 789}
 790EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 791
 792/**
 793 * nfs_request_add_commit_list - add request to a commit list
 794 * @req: pointer to a struct nfs_page
 795 * @cinfo: holds list lock and accounting info
 796 *
 797 * This sets the PG_CLEAN bit, updates the cinfo count of
 798 * number of outstanding requests requiring a commit as well as
 799 * the MM page stats.
 800 *
 801 * The caller must _not_ hold the cinfo->lock, but must be
 802 * holding the nfs_page lock.
 803 */
 804void
 805nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 806{
 807	mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 808	nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 809	mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 810	nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
 811}
 812EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 813
 814/**
 815 * nfs_request_remove_commit_list - Remove request from a commit list
 816 * @req: pointer to a nfs_page
 817 * @cinfo: holds list lock and accounting info
 818 *
 819 * This clears the PG_CLEAN bit, and updates the cinfo's count of
 820 * number of outstanding requests requiring a commit
 821 * It does not update the MM page stats.
 822 *
 823 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 824 */
 825void
 826nfs_request_remove_commit_list(struct nfs_page *req,
 827			       struct nfs_commit_info *cinfo)
 828{
 829	if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 830		return;
 831	nfs_list_remove_request(req);
 832	atomic_long_dec(&cinfo->mds->ncommit);
 833}
 834EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 835
 836static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 837				      struct inode *inode)
 838{
 839	cinfo->inode = inode;
 840	cinfo->mds = &NFS_I(inode)->commit_info;
 841	cinfo->ds = pnfs_get_ds_info(inode);
 842	cinfo->dreq = NULL;
 843	cinfo->completion_ops = &nfs_commit_completion_ops;
 844}
 845
 846void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 847		    struct inode *inode,
 848		    struct nfs_direct_req *dreq)
 849{
 850	if (dreq)
 851		nfs_init_cinfo_from_dreq(cinfo, dreq);
 852	else
 853		nfs_init_cinfo_from_inode(cinfo, inode);
 854}
 855EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 856
 857/*
 858 * Add a request to the inode's commit list.
 859 */
 860void
 861nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 862			struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 863{
 864	if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 865		return;
 866	nfs_request_add_commit_list(req, cinfo);
 867}
 868
 869static void nfs_folio_clear_commit(struct folio *folio)
 870{
 871	if (folio) {
 872		long nr = folio_nr_pages(folio);
 873
 874		node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
 875		wb_stat_mod(&inode_to_bdi(folio->mapping->host)->wb,
 876			    WB_WRITEBACK, -nr);
 877	}
 878}
 879
 880/* Called holding the request lock on @req */
 881static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
 882				     struct nfs_page *req)
 883{
 884	if (test_bit(PG_CLEAN, &req->wb_flags)) {
 885		struct nfs_open_context *ctx = nfs_req_openctx(req);
 886		struct inode *inode = d_inode(ctx->dentry);
 887
 888		mutex_lock(&NFS_I(inode)->commit_mutex);
 889		if (!pnfs_clear_request_commit(req, cinfo)) {
 890			nfs_request_remove_commit_list(req, cinfo);
 891		}
 892		mutex_unlock(&NFS_I(inode)->commit_mutex);
 893		nfs_folio_clear_commit(nfs_page_to_folio(req));
 894	}
 895}
 896
 897int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 898{
 899	if (hdr->verf.committed == NFS_DATA_SYNC)
 900		return hdr->lseg == NULL;
 901	return hdr->verf.committed != NFS_FILE_SYNC;
 902}
 903
 904static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 905{
 906	nfs_io_completion_get(hdr->io_completion);
 907}
 908
 909static void nfs_write_completion(struct nfs_pgio_header *hdr)
 910{
 911	struct nfs_commit_info cinfo;
 912	unsigned long bytes = 0;
 913
 914	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 915		goto out;
 916	nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 917	while (!list_empty(&hdr->pages)) {
 918		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 919
 920		bytes += req->wb_bytes;
 921		nfs_list_remove_request(req);
 922		if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 923		    (hdr->good_bytes < bytes)) {
 924			trace_nfs_comp_error(hdr->inode, req, hdr->error);
 925			nfs_mapping_set_error(nfs_page_to_folio(req),
 926					      hdr->error);
 927			goto remove_req;
 928		}
 929		if (nfs_write_need_commit(hdr)) {
 930			/* Reset wb_nio, since the write was successful. */
 931			req->wb_nio = 0;
 932			memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 933			nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 934				hdr->ds_commit_idx);
 935			goto next;
 936		}
 937remove_req:
 938		nfs_inode_remove_request(req);
 939next:
 940		nfs_page_end_writeback(req);
 941		nfs_release_request(req);
 942	}
 943out:
 944	nfs_io_completion_put(hdr->io_completion);
 945	hdr->release(hdr);
 946}
 947
 948unsigned long
 949nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 950{
 951	return atomic_long_read(&cinfo->mds->ncommit);
 952}
 953
 954/* NFS_I(cinfo->inode)->commit_mutex held by caller */
 955int
 956nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 957		     struct nfs_commit_info *cinfo, int max)
 958{
 959	struct nfs_page *req, *tmp;
 960	int ret = 0;
 961
 962	list_for_each_entry_safe(req, tmp, src, wb_list) {
 963		kref_get(&req->wb_kref);
 964		if (!nfs_lock_request(req)) {
 965			nfs_release_request(req);
 966			continue;
 967		}
 968		nfs_request_remove_commit_list(req, cinfo);
 969		clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
 970		nfs_list_add_request(req, dst);
 971		ret++;
 972		if ((ret == max) && !cinfo->dreq)
 973			break;
 974		cond_resched();
 975	}
 976	return ret;
 977}
 978EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
 979
 980/*
 981 * nfs_scan_commit - Scan an inode for commit requests
 982 * @inode: NFS inode to scan
 983 * @dst: mds destination list
 984 * @cinfo: mds and ds lists of reqs ready to commit
 985 *
 986 * Moves requests from the inode's 'commit' request list.
 987 * The requests are *not* checked to ensure that they form a contiguous set.
 988 */
 989int
 990nfs_scan_commit(struct inode *inode, struct list_head *dst,
 991		struct nfs_commit_info *cinfo)
 992{
 993	int ret = 0;
 994
 995	if (!atomic_long_read(&cinfo->mds->ncommit))
 996		return 0;
 997	mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 998	if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
 999		const int max = INT_MAX;
1000
1001		ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1002					   cinfo, max);
1003		ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1004	}
1005	mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1006	return ret;
1007}
1008
1009/*
1010 * Search for an existing write request, and attempt to update
1011 * it to reflect a new dirty region on a given page.
1012 *
1013 * If the attempt fails, then the existing request is flushed out
1014 * to disk.
1015 */
1016static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
1017						  unsigned int offset,
1018						  unsigned int bytes)
1019{
1020	struct nfs_page *req;
1021	unsigned int rqend;
1022	unsigned int end;
1023	int error;
1024
1025	trace_nfs_try_to_update_request(folio_inode(folio), offset, bytes);
1026	end = offset + bytes;
1027
1028	req = nfs_lock_and_join_requests(folio);
1029	if (IS_ERR_OR_NULL(req))
1030		goto out;
1031
1032	rqend = req->wb_offset + req->wb_bytes;
1033	/*
1034	 * Tell the caller to flush out the request if
1035	 * the offsets are non-contiguous.
1036	 * Note: nfs_flush_incompatible() will already
1037	 * have flushed out requests having wrong owners.
1038	 */
1039	if (offset > rqend || end < req->wb_offset)
1040		goto out_flushme;
1041
1042	/* Okay, the request matches. Update the region */
1043	if (offset < req->wb_offset) {
1044		req->wb_offset = offset;
1045		req->wb_pgbase = offset;
1046	}
1047	if (end > rqend)
1048		req->wb_bytes = end - req->wb_offset;
1049	else
1050		req->wb_bytes = rqend - req->wb_offset;
1051	req->wb_nio = 0;
1052out:
1053	trace_nfs_try_to_update_request_done(folio_inode(folio), offset, bytes,
1054					     PTR_ERR_OR_ZERO(req));
1055	return req;
1056out_flushme:
1057	/*
1058	 * Note: we mark the request dirty here because
1059	 * nfs_lock_and_join_requests() cannot preserve
1060	 * commit flags, so we have to replay the write.
1061	 */
1062	nfs_mark_request_dirty(req);
1063	nfs_unlock_and_release_request(req);
1064	error = nfs_wb_folio(folio->mapping->host, folio);
1065	trace_nfs_try_to_update_request_done(folio_inode(folio), offset, bytes, error);
1066	return (error < 0) ? ERR_PTR(error) : NULL;
1067}
1068
1069/*
1070 * Try to update an existing write request, or create one if there is none.
1071 *
1072 * Note: Should always be called with the Page Lock held to prevent races
1073 * if we have to add a new request. Also assumes that the caller has
1074 * already called nfs_flush_incompatible() if necessary.
1075 */
1076static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx,
1077						struct folio *folio,
1078						unsigned int offset,
1079						unsigned int bytes)
1080{
1081	struct nfs_page *req;
1082
1083	req = nfs_try_to_update_request(folio, offset, bytes);
1084	if (req != NULL)
1085		goto out;
1086	req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1087	if (IS_ERR(req))
1088		goto out;
1089	nfs_inode_add_request(req);
1090out:
1091	return req;
1092}
1093
1094static int nfs_writepage_setup(struct nfs_open_context *ctx,
1095			       struct folio *folio, unsigned int offset,
1096			       unsigned int count)
1097{
1098	struct nfs_page *req;
1099
1100	req = nfs_setup_write_request(ctx, folio, offset, count);
1101	if (IS_ERR(req))
1102		return PTR_ERR(req);
1103	trace_nfs_writepage_setup(req);
1104	/* Update file length */
1105	nfs_grow_file(folio, offset, count);
1106	nfs_mark_uptodate(req);
1107	nfs_mark_request_dirty(req);
1108	nfs_unlock_and_release_request(req);
1109	return 0;
1110}
1111
1112int nfs_flush_incompatible(struct file *file, struct folio *folio)
1113{
1114	struct nfs_open_context *ctx = nfs_file_open_context(file);
1115	struct nfs_lock_context *l_ctx;
1116	struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1117	struct nfs_page	*req;
1118	int do_flush, status;
1119	/*
1120	 * Look for a request corresponding to this page. If there
1121	 * is one, and it belongs to another file, we flush it out
1122	 * before we try to copy anything into the page. Do this
1123	 * due to the lack of an ACCESS-type call in NFSv2.
1124	 * Also do the same if we find a request from an existing
1125	 * dropped page.
1126	 */
1127	do {
1128		req = nfs_folio_find_head_request(folio);
1129		if (req == NULL)
1130			return 0;
1131		l_ctx = req->wb_lock_context;
1132		do_flush = nfs_page_to_folio(req) != folio ||
1133			   !nfs_match_open_context(nfs_req_openctx(req), ctx);
1134		if (l_ctx && flctx &&
1135		    !(list_empty_careful(&flctx->flc_posix) &&
1136		      list_empty_careful(&flctx->flc_flock))) {
1137			do_flush |= l_ctx->lockowner != current->files;
1138		}
1139		nfs_release_request(req);
1140		if (!do_flush)
1141			return 0;
1142		status = nfs_wb_folio(folio->mapping->host, folio);
1143	} while (status == 0);
1144	return status;
1145}
1146
1147/*
1148 * Avoid buffered writes when a open context credential's key would
1149 * expire soon.
1150 *
1151 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1152 *
1153 * Return 0 and set a credential flag which triggers the inode to flush
1154 * and performs  NFS_FILE_SYNC writes if the key will expired within
1155 * RPC_KEY_EXPIRE_TIMEO.
1156 */
1157int
1158nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1159{
1160	struct nfs_open_context *ctx = nfs_file_open_context(filp);
1161
1162	if (nfs_ctx_key_to_expire(ctx, inode) &&
1163	    !rcu_access_pointer(ctx->ll_cred))
1164		/* Already expired! */
1165		return -EACCES;
1166	return 0;
1167}
1168
1169/*
1170 * Test if the open context credential key is marked to expire soon.
1171 */
1172bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1173{
1174	struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1175	struct rpc_cred *cred, *new, *old = NULL;
1176	struct auth_cred acred = {
1177		.cred = ctx->cred,
1178	};
1179	bool ret = false;
1180
1181	rcu_read_lock();
1182	cred = rcu_dereference(ctx->ll_cred);
1183	if (cred && !(cred->cr_ops->crkey_timeout &&
1184		      cred->cr_ops->crkey_timeout(cred)))
1185		goto out;
1186	rcu_read_unlock();
1187
1188	new = auth->au_ops->lookup_cred(auth, &acred, 0);
1189	if (new == cred) {
1190		put_rpccred(new);
1191		return true;
1192	}
1193	if (IS_ERR_OR_NULL(new)) {
1194		new = NULL;
1195		ret = true;
1196	} else if (new->cr_ops->crkey_timeout &&
1197		   new->cr_ops->crkey_timeout(new))
1198		ret = true;
1199
1200	rcu_read_lock();
1201	old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1202					     RCU_INITIALIZER(new)), 1);
1203out:
1204	rcu_read_unlock();
1205	put_rpccred(old);
1206	return ret;
1207}
1208
1209/*
1210 * If the page cache is marked as unsafe or invalid, then we can't rely on
1211 * the PageUptodate() flag. In this case, we will need to turn off
1212 * write optimisations that depend on the page contents being correct.
1213 */
1214static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1215{
1216	struct inode *inode = folio->mapping->host;
1217	struct nfs_inode *nfsi = NFS_I(inode);
1218
1219	if (nfs_have_delegated_attributes(inode))
1220		goto out;
1221	if (nfsi->cache_validity &
1222	    (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1223		return false;
1224	smp_rmb();
1225	if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1226		return false;
1227out:
1228	if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1229		return false;
1230	return folio_test_uptodate(folio) != 0;
1231}
1232
1233static bool
1234is_whole_file_wrlock(struct file_lock *fl)
1235{
1236	return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1237			lock_is_write(fl);
1238}
1239
1240/* If we know the page is up to date, and we're not using byte range locks (or
1241 * if we have the whole file locked for writing), it may be more efficient to
1242 * extend the write to cover the entire page in order to avoid fragmentation
1243 * inefficiencies.
1244 *
1245 * If the file is opened for synchronous writes then we can just skip the rest
1246 * of the checks.
1247 */
1248static int nfs_can_extend_write(struct file *file, struct folio *folio,
1249				unsigned int pagelen)
1250{
1251	struct inode *inode = file_inode(file);
1252	struct file_lock_context *flctx = locks_inode_context(inode);
1253	struct file_lock *fl;
1254	int ret;
1255	unsigned int mntflags = NFS_SERVER(inode)->flags;
1256
1257	if (mntflags & NFS_MOUNT_NO_ALIGNWRITE)
1258		return 0;
1259	if (file->f_flags & O_DSYNC)
1260		return 0;
1261	if (!nfs_folio_write_uptodate(folio, pagelen))
1262		return 0;
1263	if (nfs_have_write_delegation(inode))
1264		return 1;
1265	if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1266		       list_empty_careful(&flctx->flc_posix)))
1267		return 1;
1268
1269	/* Check to see if there are whole file write locks */
1270	ret = 0;
1271	spin_lock(&flctx->flc_lock);
1272	if (!list_empty(&flctx->flc_posix)) {
1273		fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1274					c.flc_list);
1275		if (is_whole_file_wrlock(fl))
1276			ret = 1;
1277	} else if (!list_empty(&flctx->flc_flock)) {
1278		fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1279					c.flc_list);
1280		if (lock_is_write(fl))
1281			ret = 1;
1282	}
1283	spin_unlock(&flctx->flc_lock);
1284	return ret;
1285}
1286
1287/*
1288 * Update and possibly write a cached page of an NFS file.
1289 *
1290 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1291 * things with a page scheduled for an RPC call (e.g. invalidate it).
1292 */
1293int nfs_update_folio(struct file *file, struct folio *folio,
1294		     unsigned int offset, unsigned int count)
1295{
1296	struct nfs_open_context *ctx = nfs_file_open_context(file);
1297	struct address_space *mapping = folio->mapping;
1298	struct inode *inode = mapping->host;
1299	unsigned int pagelen = nfs_folio_length(folio);
1300	int		status = 0;
1301
1302	nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1303
1304	trace_nfs_update_folio(inode, offset, count);
1305
1306	dprintk("NFS:       nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1307		(long long)(folio_pos(folio) + offset));
1308
1309	if (!count)
1310		goto out;
1311
1312	if (nfs_can_extend_write(file, folio, pagelen)) {
1313		unsigned int end = count + offset;
1314
1315		offset = round_down(offset, PAGE_SIZE);
1316		if (end < pagelen)
1317			end = min(round_up(end, PAGE_SIZE), pagelen);
1318		count = end - offset;
1319	}
1320
1321	status = nfs_writepage_setup(ctx, folio, offset, count);
1322	if (status < 0)
1323		nfs_set_pageerror(mapping);
1324out:
1325	trace_nfs_update_folio_done(inode, offset, count, status);
1326	dprintk("NFS:       nfs_update_folio returns %d (isize %lld)\n",
1327			status, (long long)i_size_read(inode));
1328	return status;
1329}
1330
1331static int flush_task_priority(int how)
1332{
1333	switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1334		case FLUSH_HIGHPRI:
1335			return RPC_PRIORITY_HIGH;
1336		case FLUSH_LOWPRI:
1337			return RPC_PRIORITY_LOW;
1338	}
1339	return RPC_PRIORITY_NORMAL;
1340}
1341
1342static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1343			       struct rpc_message *msg,
1344			       const struct nfs_rpc_ops *rpc_ops,
1345			       struct rpc_task_setup *task_setup_data, int how)
1346{
1347	int priority = flush_task_priority(how);
1348
1349	if (IS_SWAPFILE(hdr->inode))
1350		task_setup_data->flags |= RPC_TASK_SWAPPER;
1351	task_setup_data->priority = priority;
1352	rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1353	trace_nfs_initiate_write(hdr);
1354}
1355
1356/* If a nfs_flush_* function fails, it should remove reqs from @head and
1357 * call this on each, which will prepare them to be retried on next
1358 * writeback using standard nfs.
1359 */
1360static void nfs_redirty_request(struct nfs_page *req)
1361{
1362	struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1363
1364	/* Bump the transmission count */
1365	req->wb_nio++;
1366	nfs_mark_request_dirty(req);
1367	atomic_long_inc(&nfsi->redirtied_pages);
1368	nfs_page_end_writeback(req);
1369	nfs_release_request(req);
1370}
1371
1372static void nfs_async_write_error(struct list_head *head, int error)
1373{
1374	struct nfs_page	*req;
1375
1376	while (!list_empty(head)) {
1377		req = nfs_list_entry(head->next);
1378		nfs_list_remove_request(req);
1379		if (nfs_error_is_fatal_on_server(error))
1380			nfs_write_error(req, error);
1381		else
1382			nfs_redirty_request(req);
1383	}
1384}
1385
1386static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1387{
1388	nfs_async_write_error(&hdr->pages, 0);
1389}
1390
1391static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1392	.init_hdr = nfs_async_write_init,
1393	.error_cleanup = nfs_async_write_error,
1394	.completion = nfs_write_completion,
1395	.reschedule_io = nfs_async_write_reschedule_io,
1396};
1397
1398void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1399			       struct inode *inode, int ioflags, bool force_mds,
1400			       const struct nfs_pgio_completion_ops *compl_ops)
1401{
1402	struct nfs_server *server = NFS_SERVER(inode);
1403	const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1404
1405#ifdef CONFIG_NFS_V4_1
1406	if (server->pnfs_curr_ld && !force_mds)
1407		pg_ops = server->pnfs_curr_ld->pg_write_ops;
1408#endif
1409	nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1410			server->wsize, ioflags);
1411}
1412EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1413
1414void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1415{
1416	struct nfs_pgio_mirror *mirror;
1417
1418	if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1419		pgio->pg_ops->pg_cleanup(pgio);
1420
1421	pgio->pg_ops = &nfs_pgio_rw_ops;
1422
1423	nfs_pageio_stop_mirroring(pgio);
1424
1425	mirror = &pgio->pg_mirrors[0];
1426	mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1427}
1428EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1429
1430
1431void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1432{
1433	struct nfs_commit_data *data = calldata;
1434
1435	NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1436}
1437
1438static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1439		struct nfs_fattr *fattr)
1440{
1441	struct nfs_pgio_args *argp = &hdr->args;
1442	struct nfs_pgio_res *resp = &hdr->res;
1443	u64 size = argp->offset + resp->count;
1444
1445	if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1446		fattr->size = size;
1447	if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1448		fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1449		return;
1450	}
1451	if (size != fattr->size)
1452		return;
1453	/* Set attribute barrier */
1454	nfs_fattr_set_barrier(fattr);
1455	/* ...and update size */
1456	fattr->valid |= NFS_ATTR_FATTR_SIZE;
1457}
1458
1459void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1460{
1461	struct nfs_fattr *fattr = &hdr->fattr;
1462	struct inode *inode = hdr->inode;
1463
1464	if (nfs_have_delegated_mtime(inode)) {
1465		spin_lock(&inode->i_lock);
1466		nfs_set_cache_invalid(inode, NFS_INO_INVALID_BLOCKS);
1467		spin_unlock(&inode->i_lock);
1468		return;
1469	}
1470
1471	spin_lock(&inode->i_lock);
1472	nfs_writeback_check_extend(hdr, fattr);
1473	nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1474	spin_unlock(&inode->i_lock);
1475}
1476EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1477
1478/*
1479 * This function is called when the WRITE call is complete.
1480 */
1481static int nfs_writeback_done(struct rpc_task *task,
1482			      struct nfs_pgio_header *hdr,
1483			      struct inode *inode)
1484{
1485	int status;
1486
1487	/*
1488	 * ->write_done will attempt to use post-op attributes to detect
1489	 * conflicting writes by other clients.  A strict interpretation
1490	 * of close-to-open would allow us to continue caching even if
1491	 * another writer had changed the file, but some applications
1492	 * depend on tighter cache coherency when writing.
1493	 */
1494	status = NFS_PROTO(inode)->write_done(task, hdr);
1495	if (status != 0)
1496		return status;
1497
1498	nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1499	trace_nfs_writeback_done(task, hdr);
1500
1501	if (task->tk_status >= 0) {
1502		enum nfs3_stable_how committed = hdr->res.verf->committed;
1503
1504		if (committed == NFS_UNSTABLE) {
1505			/*
1506			 * We have some uncommitted data on the server at
1507			 * this point, so ensure that we keep track of that
1508			 * fact irrespective of what later writes do.
1509			 */
1510			set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1511		}
1512
1513		if (committed < hdr->args.stable) {
1514			/* We tried a write call, but the server did not
1515			 * commit data to stable storage even though we
1516			 * requested it.
1517			 * Note: There is a known bug in Tru64 < 5.0 in which
1518			 *	 the server reports NFS_DATA_SYNC, but performs
1519			 *	 NFS_FILE_SYNC. We therefore implement this checking
1520			 *	 as a dprintk() in order to avoid filling syslog.
1521			 */
1522			static unsigned long    complain;
1523
1524			/* Note this will print the MDS for a DS write */
1525			if (time_before(complain, jiffies)) {
1526				dprintk("NFS:       faulty NFS server %s:"
1527					" (committed = %d) != (stable = %d)\n",
1528					NFS_SERVER(inode)->nfs_client->cl_hostname,
1529					committed, hdr->args.stable);
1530				complain = jiffies + 300 * HZ;
1531			}
1532		}
1533	}
1534
1535	/* Deal with the suid/sgid bit corner case */
1536	if (nfs_should_remove_suid(inode)) {
1537		spin_lock(&inode->i_lock);
1538		nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE
1539				| NFS_INO_REVAL_FORCED);
1540		spin_unlock(&inode->i_lock);
1541	}
1542	return 0;
1543}
1544
1545/*
1546 * This function is called when the WRITE call is complete.
1547 */
1548static void nfs_writeback_result(struct rpc_task *task,
1549				 struct nfs_pgio_header *hdr)
1550{
1551	struct nfs_pgio_args	*argp = &hdr->args;
1552	struct nfs_pgio_res	*resp = &hdr->res;
1553
1554	if (resp->count < argp->count) {
1555		static unsigned long    complain;
1556
1557		/* This a short write! */
1558		nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1559
1560		/* Has the server at least made some progress? */
1561		if (resp->count == 0) {
1562			if (time_before(complain, jiffies)) {
1563				printk(KERN_WARNING
1564				       "NFS: Server wrote zero bytes, expected %u.\n",
1565				       argp->count);
1566				complain = jiffies + 300 * HZ;
1567			}
1568			nfs_set_pgio_error(hdr, -EIO, argp->offset);
1569			task->tk_status = -EIO;
1570			return;
1571		}
1572
1573		/* For non rpc-based layout drivers, retry-through-MDS */
1574		if (!task->tk_ops) {
1575			hdr->pnfs_error = -EAGAIN;
1576			return;
1577		}
1578
1579		/* Was this an NFSv2 write or an NFSv3 stable write? */
1580		if (resp->verf->committed != NFS_UNSTABLE) {
1581			/* Resend from where the server left off */
1582			hdr->mds_offset += resp->count;
1583			argp->offset += resp->count;
1584			argp->pgbase += resp->count;
1585			argp->count -= resp->count;
1586		} else {
1587			/* Resend as a stable write in order to avoid
1588			 * headaches in the case of a server crash.
1589			 */
1590			argp->stable = NFS_FILE_SYNC;
1591		}
1592		resp->count = 0;
1593		resp->verf->committed = 0;
1594		rpc_restart_call_prepare(task);
1595	}
1596}
1597
1598static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1599{
1600	return wait_var_event_killable(&cinfo->rpcs_out,
1601				       !atomic_read(&cinfo->rpcs_out));
1602}
1603
1604void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1605{
1606	atomic_inc(&cinfo->rpcs_out);
1607}
1608
1609bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1610{
1611	if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1612		wake_up_var(&cinfo->rpcs_out);
1613		return true;
1614	}
1615	return false;
1616}
1617
1618void nfs_commitdata_release(struct nfs_commit_data *data)
1619{
1620	put_nfs_open_context(data->context);
1621	nfs_commit_free(data);
1622}
1623EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1624
1625int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1626			const struct nfs_rpc_ops *nfs_ops,
1627			const struct rpc_call_ops *call_ops,
1628			int how, int flags,
1629			struct nfsd_file *localio)
1630{
1631	struct rpc_task *task;
1632	int priority = flush_task_priority(how);
1633	struct rpc_message msg = {
1634		.rpc_argp = &data->args,
1635		.rpc_resp = &data->res,
1636		.rpc_cred = data->cred,
1637	};
1638	struct rpc_task_setup task_setup_data = {
1639		.task = &data->task,
1640		.rpc_client = clnt,
1641		.rpc_message = &msg,
1642		.callback_ops = call_ops,
1643		.callback_data = data,
1644		.workqueue = nfsiod_workqueue,
1645		.flags = RPC_TASK_ASYNC | flags,
1646		.priority = priority,
1647	};
1648
1649	if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1650		task_setup_data.flags |= RPC_TASK_MOVEABLE;
1651
1652	/* Set up the initial task struct.  */
1653	nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1654	trace_nfs_initiate_commit(data);
1655
1656	dprintk("NFS: initiated commit call\n");
1657
1658	if (localio)
1659		return nfs_local_commit(localio, data, call_ops, how);
1660
1661	task = rpc_run_task(&task_setup_data);
1662	if (IS_ERR(task))
1663		return PTR_ERR(task);
1664	if (how & FLUSH_SYNC)
1665		rpc_wait_for_completion_task(task);
1666	rpc_put_task(task);
1667	return 0;
1668}
1669EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1670
1671static loff_t nfs_get_lwb(struct list_head *head)
1672{
1673	loff_t lwb = 0;
1674	struct nfs_page *req;
1675
1676	list_for_each_entry(req, head, wb_list)
1677		if (lwb < (req_offset(req) + req->wb_bytes))
1678			lwb = req_offset(req) + req->wb_bytes;
1679
1680	return lwb;
1681}
1682
1683/*
1684 * Set up the argument/result storage required for the RPC call.
1685 */
1686void nfs_init_commit(struct nfs_commit_data *data,
1687		     struct list_head *head,
1688		     struct pnfs_layout_segment *lseg,
1689		     struct nfs_commit_info *cinfo)
1690{
1691	struct nfs_page *first;
1692	struct nfs_open_context *ctx;
1693	struct inode *inode;
1694
1695	/* Set up the RPC argument and reply structs
1696	 * NB: take care not to mess about with data->commit et al. */
1697
1698	if (head)
1699		list_splice_init(head, &data->pages);
1700
1701	first = nfs_list_entry(data->pages.next);
1702	ctx = nfs_req_openctx(first);
1703	inode = d_inode(ctx->dentry);
1704
1705	data->inode	  = inode;
1706	data->cred	  = ctx->cred;
1707	data->lseg	  = lseg; /* reference transferred */
1708	/* only set lwb for pnfs commit */
1709	if (lseg)
1710		data->lwb = nfs_get_lwb(&data->pages);
1711	data->mds_ops     = &nfs_commit_ops;
1712	data->completion_ops = cinfo->completion_ops;
1713	data->dreq	  = cinfo->dreq;
1714
1715	data->args.fh     = NFS_FH(data->inode);
1716	/* Note: we always request a commit of the entire inode */
1717	data->args.offset = 0;
1718	data->args.count  = 0;
1719	data->context     = get_nfs_open_context(ctx);
1720	data->res.fattr   = &data->fattr;
1721	data->res.verf    = &data->verf;
1722	nfs_fattr_init(&data->fattr);
1723	nfs_commit_begin(cinfo->mds);
1724}
1725EXPORT_SYMBOL_GPL(nfs_init_commit);
1726
1727void nfs_retry_commit(struct list_head *page_list,
1728		      struct pnfs_layout_segment *lseg,
1729		      struct nfs_commit_info *cinfo,
1730		      u32 ds_commit_idx)
1731{
1732	struct nfs_page *req;
1733
1734	while (!list_empty(page_list)) {
1735		req = nfs_list_entry(page_list->next);
1736		nfs_list_remove_request(req);
1737		nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1738		nfs_folio_clear_commit(nfs_page_to_folio(req));
1739		nfs_unlock_and_release_request(req);
1740	}
1741}
1742EXPORT_SYMBOL_GPL(nfs_retry_commit);
1743
1744static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1745				     struct nfs_page *req)
1746{
1747	struct folio *folio = nfs_page_to_folio(req);
1748
1749	filemap_dirty_folio(folio_mapping(folio), folio);
1750}
1751
1752/*
1753 * Commit dirty pages
1754 */
1755static int
1756nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1757		struct nfs_commit_info *cinfo)
1758{
1759	struct nfs_commit_data	*data;
1760	struct nfsd_file *localio;
1761	unsigned short task_flags = 0;
1762
1763	/* another commit raced with us */
1764	if (list_empty(head))
1765		return 0;
1766
1767	data = nfs_commitdata_alloc();
1768	if (!data) {
1769		nfs_retry_commit(head, NULL, cinfo, -1);
1770		return -ENOMEM;
1771	}
1772
1773	/* Set up the argument struct */
1774	nfs_init_commit(data, head, NULL, cinfo);
1775	if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1776		task_flags = RPC_TASK_MOVEABLE;
1777
1778	localio = nfs_local_open_fh(NFS_SERVER(inode)->nfs_client, data->cred,
1779				    data->args.fh, &data->context->nfl,
1780				    data->context->mode);
1781	return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1782				   data->mds_ops, how,
1783				   RPC_TASK_CRED_NOREF | task_flags, localio);
1784}
1785
1786/*
1787 * COMMIT call returned
1788 */
1789static void nfs_commit_done(struct rpc_task *task, void *calldata)
1790{
1791	struct nfs_commit_data	*data = calldata;
1792
1793	/* Call the NFS version-specific code */
1794	NFS_PROTO(data->inode)->commit_done(task, data);
1795	trace_nfs_commit_done(task, data);
1796}
1797
1798static void nfs_commit_release_pages(struct nfs_commit_data *data)
1799{
1800	const struct nfs_writeverf *verf = data->res.verf;
1801	struct nfs_page	*req;
1802	int status = data->task.tk_status;
1803	struct nfs_commit_info cinfo;
1804	struct folio *folio;
1805
1806	while (!list_empty(&data->pages)) {
1807		req = nfs_list_entry(data->pages.next);
1808		nfs_list_remove_request(req);
1809		folio = nfs_page_to_folio(req);
1810		nfs_folio_clear_commit(folio);
1811
1812		dprintk("NFS:       commit (%s/%llu %d@%lld)",
1813			nfs_req_openctx(req)->dentry->d_sb->s_id,
1814			(unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1815			req->wb_bytes,
1816			(long long)req_offset(req));
1817		if (status < 0) {
1818			if (folio) {
1819				trace_nfs_commit_error(data->inode, req,
1820						       status);
1821				nfs_mapping_set_error(folio, status);
1822				nfs_inode_remove_request(req);
1823			}
1824			dprintk(", error = %d\n", status);
1825			goto next;
1826		}
1827
1828		/* Okay, COMMIT succeeded, apparently. Check the verifier
1829		 * returned by the server against all stored verfs. */
1830		if (nfs_write_match_verf(verf, req)) {
1831			/* We have a match */
1832			if (folio)
1833				nfs_inode_remove_request(req);
1834			dprintk(" OK\n");
1835			goto next;
1836		}
1837		/* We have a mismatch. Write the page again */
1838		dprintk(" mismatch\n");
1839		nfs_mark_request_dirty(req);
1840		atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1841	next:
1842		nfs_unlock_and_release_request(req);
1843		/* Latency breaker */
1844		cond_resched();
1845	}
1846
1847	nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1848	nfs_commit_end(cinfo.mds);
1849}
1850
1851static void nfs_commit_release(void *calldata)
1852{
1853	struct nfs_commit_data *data = calldata;
1854
1855	data->completion_ops->completion(data);
1856	nfs_commitdata_release(calldata);
1857}
1858
1859static const struct rpc_call_ops nfs_commit_ops = {
1860	.rpc_call_prepare = nfs_commit_prepare,
1861	.rpc_call_done = nfs_commit_done,
1862	.rpc_release = nfs_commit_release,
1863};
1864
1865static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1866	.completion = nfs_commit_release_pages,
1867	.resched_write = nfs_commit_resched_write,
1868};
1869
1870int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1871			    int how, struct nfs_commit_info *cinfo)
1872{
1873	int status;
1874
1875	status = pnfs_commit_list(inode, head, how, cinfo);
1876	if (status == PNFS_NOT_ATTEMPTED)
1877		status = nfs_commit_list(inode, head, how, cinfo);
1878	return status;
1879}
1880
1881static int __nfs_commit_inode(struct inode *inode, int how,
1882		struct writeback_control *wbc)
1883{
1884	LIST_HEAD(head);
1885	struct nfs_commit_info cinfo;
1886	int may_wait = how & FLUSH_SYNC;
1887	int ret, nscan;
1888
1889	how &= ~FLUSH_SYNC;
1890	nfs_init_cinfo_from_inode(&cinfo, inode);
1891	nfs_commit_begin(cinfo.mds);
1892	for (;;) {
1893		ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1894		if (ret <= 0)
1895			break;
1896		ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1897		if (ret < 0)
1898			break;
1899		ret = 0;
1900		if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1901			if (nscan < wbc->nr_to_write)
1902				wbc->nr_to_write -= nscan;
1903			else
1904				wbc->nr_to_write = 0;
1905		}
1906		if (nscan < INT_MAX)
1907			break;
1908		cond_resched();
1909	}
1910	nfs_commit_end(cinfo.mds);
1911	if (ret || !may_wait)
1912		return ret;
1913	return wait_on_commit(cinfo.mds);
1914}
1915
1916int nfs_commit_inode(struct inode *inode, int how)
1917{
1918	return __nfs_commit_inode(inode, how, NULL);
1919}
1920EXPORT_SYMBOL_GPL(nfs_commit_inode);
1921
1922int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1923{
1924	struct nfs_inode *nfsi = NFS_I(inode);
1925	int flags = FLUSH_SYNC;
1926	int ret = 0;
1927
1928	if (wbc->sync_mode == WB_SYNC_NONE) {
1929		/* no commits means nothing needs to be done */
1930		if (!atomic_long_read(&nfsi->commit_info.ncommit))
1931			goto check_requests_outstanding;
1932
1933		/* Don't commit yet if this is a non-blocking flush and there
1934		 * are a lot of outstanding writes for this mapping.
1935		 */
1936		if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1937			goto out_mark_dirty;
1938
1939		/* don't wait for the COMMIT response */
1940		flags = 0;
1941	}
1942
1943	ret = __nfs_commit_inode(inode, flags, wbc);
1944	if (!ret) {
1945		if (flags & FLUSH_SYNC)
1946			return 0;
1947	} else if (atomic_long_read(&nfsi->commit_info.ncommit))
1948		goto out_mark_dirty;
1949
1950check_requests_outstanding:
1951	if (!atomic_read(&nfsi->commit_info.rpcs_out))
1952		return ret;
1953out_mark_dirty:
1954	__mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1955	return ret;
1956}
1957EXPORT_SYMBOL_GPL(nfs_write_inode);
1958
1959/*
1960 * Wrapper for filemap_write_and_wait_range()
1961 *
1962 * Needed for pNFS in order to ensure data becomes visible to the
1963 * client.
1964 */
1965int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1966		loff_t lstart, loff_t lend)
1967{
1968	int ret;
1969
1970	ret = filemap_write_and_wait_range(mapping, lstart, lend);
1971	if (ret == 0)
1972		ret = pnfs_sync_inode(mapping->host, true);
1973	return ret;
1974}
1975EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
1976
1977/*
1978 * flush the inode to disk.
1979 */
1980int nfs_wb_all(struct inode *inode)
1981{
1982	int ret;
1983
1984	trace_nfs_writeback_inode_enter(inode);
1985
1986	ret = filemap_write_and_wait(inode->i_mapping);
1987	if (ret)
1988		goto out;
1989	ret = nfs_commit_inode(inode, FLUSH_SYNC);
1990	if (ret < 0)
1991		goto out;
1992	pnfs_sync_inode(inode, true);
1993	ret = 0;
1994
1995out:
1996	trace_nfs_writeback_inode_exit(inode, ret);
1997	return ret;
1998}
1999EXPORT_SYMBOL_GPL(nfs_wb_all);
2000
2001int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2002{
2003	struct nfs_page *req;
2004	int ret = 0;
2005
2006	folio_wait_writeback(folio);
2007
2008	/* blocking call to cancel all requests and join to a single (head)
2009	 * request */
2010	req = nfs_lock_and_join_requests(folio);
2011
2012	if (IS_ERR(req)) {
2013		ret = PTR_ERR(req);
2014	} else if (req) {
2015		/* all requests from this folio have been cancelled by
2016		 * nfs_lock_and_join_requests, so just remove the head
2017		 * request from the inode / page_private pointer and
2018		 * release it */
2019		nfs_inode_remove_request(req);
2020		nfs_unlock_and_release_request(req);
2021		folio_cancel_dirty(folio);
2022	}
2023
2024	return ret;
2025}
2026
2027/**
2028 * nfs_wb_folio_reclaim - Write back all requests on one page
2029 * @inode: pointer to page
2030 * @folio: pointer to folio
2031 *
2032 * Assumes that the folio has been locked by the caller
2033 */
2034int nfs_wb_folio_reclaim(struct inode *inode, struct folio *folio)
2035{
2036	loff_t range_start = folio_pos(folio);
2037	size_t len = folio_size(folio);
2038	struct writeback_control wbc = {
2039		.sync_mode = WB_SYNC_ALL,
2040		.nr_to_write = 0,
2041		.range_start = range_start,
2042		.range_end = range_start + len - 1,
2043		.for_sync = 1,
2044	};
2045	int ret;
2046
2047	if (folio_test_writeback(folio))
2048		return -EBUSY;
2049	if (folio_clear_dirty_for_io(folio)) {
2050		trace_nfs_writeback_folio_reclaim(inode, range_start, len);
2051		ret = nfs_writepage_locked(folio, &wbc);
2052		trace_nfs_writeback_folio_reclaim_done(inode, range_start, len,
2053						       ret);
2054		return ret;
2055	}
2056	nfs_commit_inode(inode, 0);
2057	return 0;
2058}
2059
2060/**
2061 * nfs_wb_folio - Write back all requests on one page
2062 * @inode: pointer to page
2063 * @folio: pointer to folio
2064 *
2065 * Assumes that the folio has been locked by the caller, and will
2066 * not unlock it.
2067 */
2068int nfs_wb_folio(struct inode *inode, struct folio *folio)
2069{
2070	loff_t range_start = folio_pos(folio);
2071	size_t len = folio_size(folio);
2072	struct writeback_control wbc = {
2073		.sync_mode = WB_SYNC_ALL,
2074		.nr_to_write = 0,
2075		.range_start = range_start,
2076		.range_end = range_start + len - 1,
2077	};
2078	int ret;
2079
2080	trace_nfs_writeback_folio(inode, range_start, len);
2081
2082	for (;;) {
2083		folio_wait_writeback(folio);
2084		if (folio_clear_dirty_for_io(folio)) {
2085			ret = nfs_writepage_locked(folio, &wbc);
2086			if (ret < 0)
2087				goto out_error;
2088			continue;
2089		}
2090		ret = 0;
2091		if (!folio_test_private(folio))
2092			break;
2093		ret = nfs_commit_inode(inode, FLUSH_SYNC);
2094		if (ret < 0)
2095			goto out_error;
2096	}
2097out_error:
2098	trace_nfs_writeback_folio_done(inode, range_start, len, ret);
2099	return ret;
2100}
2101
2102#ifdef CONFIG_MIGRATION
2103int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2104		struct folio *src, enum migrate_mode mode)
2105{
2106	/*
2107	 * If the private flag is set, the folio is currently associated with
2108	 * an in-progress read or write request. Don't try to migrate it.
2109	 *
2110	 * FIXME: we could do this in principle, but we'll need a way to ensure
2111	 *        that we can safely release the inode reference while holding
2112	 *        the folio lock.
2113	 */
2114	if (folio_test_private(src)) {
2115		if (mode == MIGRATE_SYNC)
2116			nfs_wb_folio(src->mapping->host, src);
2117		if (folio_test_private(src))
2118			return -EBUSY;
2119	}
2120
2121	if (folio_test_private_2(src)) { /* [DEPRECATED] */
2122		if (mode == MIGRATE_ASYNC)
2123			return -EBUSY;
2124		folio_wait_private_2(src);
2125	}
2126
2127	return migrate_folio(mapping, dst, src, mode);
2128}
2129#endif
2130
2131int __init nfs_init_writepagecache(void)
2132{
2133	nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2134					     sizeof(struct nfs_pgio_header),
2135					     0, SLAB_HWCACHE_ALIGN,
2136					     NULL);
2137	if (nfs_wdata_cachep == NULL)
2138		return -ENOMEM;
2139
2140	nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2141						     nfs_wdata_cachep);
2142	if (nfs_wdata_mempool == NULL)
2143		goto out_destroy_write_cache;
2144
2145	nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2146					     sizeof(struct nfs_commit_data),
2147					     0, SLAB_HWCACHE_ALIGN,
2148					     NULL);
2149	if (nfs_cdata_cachep == NULL)
2150		goto out_destroy_write_mempool;
2151
2152	nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2153						      nfs_cdata_cachep);
2154	if (nfs_commit_mempool == NULL)
2155		goto out_destroy_commit_cache;
2156
2157	/*
2158	 * NFS congestion size, scale with available memory.
2159	 *
2160	 *  64MB:    8192k
2161	 * 128MB:   11585k
2162	 * 256MB:   16384k
2163	 * 512MB:   23170k
2164	 *   1GB:   32768k
2165	 *   2GB:   46340k
2166	 *   4GB:   65536k
2167	 *   8GB:   92681k
2168	 *  16GB:  131072k
2169	 *
2170	 * This allows larger machines to have larger/more transfers.
2171	 * Limit the default to 256M
2172	 */
2173	nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2174	if (nfs_congestion_kb > 256*1024)
2175		nfs_congestion_kb = 256*1024;
2176
2177	return 0;
2178
2179out_destroy_commit_cache:
2180	kmem_cache_destroy(nfs_cdata_cachep);
2181out_destroy_write_mempool:
2182	mempool_destroy(nfs_wdata_mempool);
2183out_destroy_write_cache:
2184	kmem_cache_destroy(nfs_wdata_cachep);
2185	return -ENOMEM;
2186}
2187
2188void nfs_destroy_writepagecache(void)
2189{
2190	mempool_destroy(nfs_commit_mempool);
2191	kmem_cache_destroy(nfs_cdata_cachep);
2192	mempool_destroy(nfs_wdata_mempool);
2193	kmem_cache_destroy(nfs_wdata_cachep);
2194}
2195
2196static const struct nfs_rw_ops nfs_rw_write_ops = {
2197	.rw_alloc_header	= nfs_writehdr_alloc,
2198	.rw_free_header		= nfs_writehdr_free,
2199	.rw_done		= nfs_writeback_done,
2200	.rw_result		= nfs_writeback_result,
2201	.rw_initiate		= nfs_initiate_write,
2202};