Linux Audio

Check our new training course

Loading...
v4.17
 
   1/*
   2 * linux/fs/nfs/direct.c
   3 *
   4 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5 *
   6 * High-performance uncached I/O for the Linux NFS client
   7 *
   8 * There are important applications whose performance or correctness
   9 * depends on uncached access to file data.  Database clusters
  10 * (multiple copies of the same instance running on separate hosts)
  11 * implement their own cache coherency protocol that subsumes file
  12 * system cache protocols.  Applications that process datasets
  13 * considerably larger than the client's memory do not always benefit
  14 * from a local cache.  A streaming video server, for instance, has no
  15 * need to cache the contents of a file.
  16 *
  17 * When an application requests uncached I/O, all read and write requests
  18 * are made directly to the server; data stored or fetched via these
  19 * requests is not cached in the Linux page cache.  The client does not
  20 * correct unaligned requests from applications.  All requested bytes are
  21 * held on permanent storage before a direct write system call returns to
  22 * an application.
  23 *
  24 * Solaris implements an uncached I/O facility called directio() that
  25 * is used for backups and sequential I/O to very large files.  Solaris
  26 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27 * an undocumented mount option.
  28 *
  29 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30 * help from Andrew Morton.
  31 *
  32 * 18 Dec 2001	Initial implementation for 2.4  --cel
  33 * 08 Jul 2002	Version for 2.4.19, with bug fixes --trondmy
  34 * 08 Jun 2003	Port to 2.5 APIs  --cel
  35 * 31 Mar 2004	Handle direct I/O without VFS support  --cel
  36 * 15 Sep 2004	Parallel async reads  --cel
  37 * 04 May 2005	support O_DIRECT with aio  --cel
  38 *
  39 */
  40
  41#include <linux/errno.h>
  42#include <linux/sched.h>
  43#include <linux/kernel.h>
  44#include <linux/file.h>
  45#include <linux/pagemap.h>
  46#include <linux/kref.h>
  47#include <linux/slab.h>
  48#include <linux/task_io_accounting_ops.h>
  49#include <linux/module.h>
  50
  51#include <linux/nfs_fs.h>
  52#include <linux/nfs_page.h>
  53#include <linux/sunrpc/clnt.h>
  54
  55#include <linux/uaccess.h>
  56#include <linux/atomic.h>
  57
  58#include "internal.h"
  59#include "iostat.h"
  60#include "pnfs.h"
 
 
  61
  62#define NFSDBG_FACILITY		NFSDBG_VFS
  63
  64static struct kmem_cache *nfs_direct_cachep;
  65
  66/*
  67 * This represents a set of asynchronous requests that we're waiting on
  68 */
  69struct nfs_direct_mirror {
  70	ssize_t count;
  71};
  72
  73struct nfs_direct_req {
  74	struct kref		kref;		/* release manager */
  75
  76	/* I/O parameters */
  77	struct nfs_open_context	*ctx;		/* file open context info */
  78	struct nfs_lock_context *l_ctx;		/* Lock context info */
  79	struct kiocb *		iocb;		/* controlling i/o request */
  80	struct inode *		inode;		/* target file of i/o */
  81
  82	/* completion state */
  83	atomic_t		io_count;	/* i/os we're waiting for */
  84	spinlock_t		lock;		/* protect completion state */
  85
  86	struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
  87	int			mirror_count;
  88
  89	loff_t			io_start;	/* Start offset for I/O */
  90	ssize_t			count,		/* bytes actually processed */
  91				max_count,	/* max expected count */
  92				bytes_left,	/* bytes left to be sent */
  93				error;		/* any reported error */
  94	struct completion	completion;	/* wait for i/o completion */
  95
  96	/* commit state */
  97	struct nfs_mds_commit_info mds_cinfo;	/* Storage for cinfo */
  98	struct pnfs_ds_commit_info ds_cinfo;	/* Storage for cinfo */
  99	struct work_struct	work;
 100	int			flags;
 101#define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
 102#define NFS_ODIRECT_RESCHED_WRITES	(2)	/* write verification failed */
 103	struct nfs_writeverf	verf;		/* unstable write verifier */
 104};
 105
 106static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 107static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 108static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
 109static void nfs_direct_write_schedule_work(struct work_struct *work);
 110
 111static inline void get_dreq(struct nfs_direct_req *dreq)
 112{
 113	atomic_inc(&dreq->io_count);
 114}
 115
 116static inline int put_dreq(struct nfs_direct_req *dreq)
 117{
 118	return atomic_dec_and_test(&dreq->io_count);
 119}
 120
 121static void
 122nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
 123{
 124	int i;
 125	ssize_t count;
 126
 127	WARN_ON_ONCE(dreq->count >= dreq->max_count);
 128
 129	if (dreq->mirror_count == 1) {
 130		dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
 131		dreq->count += hdr->good_bytes;
 132	} else {
 133		/* mirrored writes */
 134		count = dreq->mirrors[hdr->pgio_mirror_idx].count;
 135		if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
 136			count = hdr->io_start + hdr->good_bytes - dreq->io_start;
 137			dreq->mirrors[hdr->pgio_mirror_idx].count = count;
 138		}
 139		/* update the dreq->count by finding the minimum agreed count from all
 140		 * mirrors */
 141		count = dreq->mirrors[0].count;
 142
 143		for (i = 1; i < dreq->mirror_count; i++)
 144			count = min(count, dreq->mirrors[i].count);
 145
 146		dreq->count = count;
 147	}
 148}
 149
 150/*
 151 * nfs_direct_select_verf - select the right verifier
 152 * @dreq - direct request possibly spanning multiple servers
 153 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 154 * @commit_idx - commit bucket index for the DS
 155 *
 156 * returns the correct verifier to use given the role of the server
 157 */
 158static struct nfs_writeverf *
 159nfs_direct_select_verf(struct nfs_direct_req *dreq,
 160		       struct nfs_client *ds_clp,
 161		       int commit_idx)
 162{
 163	struct nfs_writeverf *verfp = &dreq->verf;
 164
 165#ifdef CONFIG_NFS_V4_1
 166	/*
 167	 * pNFS is in use, use the DS verf except commit_through_mds is set
 168	 * for layout segment where nbuckets is zero.
 169	 */
 170	if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 171		if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 172			verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 173		else
 174			WARN_ON_ONCE(1);
 175	}
 176#endif
 177	return verfp;
 178}
 179
 180
 181/*
 182 * nfs_direct_set_hdr_verf - set the write/commit verifier
 183 * @dreq - direct request possibly spanning multiple servers
 184 * @hdr - pageio header to validate against previously seen verfs
 185 *
 186 * Set the server's (MDS or DS) "seen" verifier
 187 */
 188static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 189				    struct nfs_pgio_header *hdr)
 190{
 191	struct nfs_writeverf *verfp;
 
 192
 193	verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 194	WARN_ON_ONCE(verfp->committed >= 0);
 195	memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 196	WARN_ON_ONCE(verfp->committed < 0);
 197}
 198
 199static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
 200		const struct nfs_writeverf *v2)
 201{
 202	return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
 203}
 204
 205/*
 206 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 207 * @dreq - direct request possibly spanning multiple servers
 208 * @hdr - pageio header to validate against previously seen verf
 209 *
 210 * set the server's "seen" verf if not initialized.
 211 * returns result of comparison between @hdr->verf and the "seen"
 212 * verf of the server used by @hdr (DS or MDS)
 213 */
 214static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 215					  struct nfs_pgio_header *hdr)
 216{
 217	struct nfs_writeverf *verfp;
 218
 219	verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 220	if (verfp->committed < 0) {
 221		nfs_direct_set_hdr_verf(dreq, hdr);
 222		return 0;
 223	}
 224	return nfs_direct_cmp_verf(verfp, &hdr->verf);
 225}
 226
 227/*
 228 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 229 * @dreq - direct request possibly spanning multiple servers
 230 * @data - commit data to validate against previously seen verf
 231 *
 232 * returns result of comparison between @data->verf and the verf of
 233 * the server used by @data (DS or MDS)
 234 */
 235static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 236					   struct nfs_commit_data *data)
 237{
 238	struct nfs_writeverf *verfp;
 
 239
 240	verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 241					 data->ds_commit_index);
 242
 243	/* verifier not set so always fail */
 244	if (verfp->committed < 0)
 245		return 1;
 246
 247	return nfs_direct_cmp_verf(verfp, &data->verf);
 248}
 249
 250/**
 251 * nfs_direct_IO - NFS address space operation for direct I/O
 252 * @iocb: target I/O control block
 253 * @iter: I/O buffer
 254 *
 255 * The presence of this routine in the address space ops vector means
 256 * the NFS client supports direct I/O. However, for most direct IO, we
 257 * shunt off direct read and write requests before the VFS gets them,
 258 * so this method is only ever called for swap.
 259 */
 260ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 261{
 262	struct inode *inode = iocb->ki_filp->f_mapping->host;
 263
 264	/* we only support swap file calling nfs_direct_IO */
 265	if (!IS_SWAPFILE(inode))
 266		return 0;
 267
 268	VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 269
 270	if (iov_iter_rw(iter) == READ)
 271		return nfs_file_direct_read(iocb, iter);
 272	return nfs_file_direct_write(iocb, iter);
 
 
 
 
 273}
 274
 275static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 276{
 277	unsigned int i;
 278	for (i = 0; i < npages; i++)
 279		put_page(pages[i]);
 280}
 281
 282void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 283			      struct nfs_direct_req *dreq)
 284{
 285	cinfo->inode = dreq->inode;
 286	cinfo->mds = &dreq->mds_cinfo;
 287	cinfo->ds = &dreq->ds_cinfo;
 288	cinfo->dreq = dreq;
 289	cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 290}
 291
 292static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 293					     struct nfs_pageio_descriptor *pgio,
 294					     struct nfs_page *req)
 295{
 296	int mirror_count = 1;
 297
 298	if (pgio->pg_ops->pg_get_mirror_count)
 299		mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 300
 301	dreq->mirror_count = mirror_count;
 302}
 303
 304static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 305{
 306	struct nfs_direct_req *dreq;
 307
 308	dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 309	if (!dreq)
 310		return NULL;
 311
 312	kref_init(&dreq->kref);
 313	kref_get(&dreq->kref);
 314	init_completion(&dreq->completion);
 315	INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 316	dreq->verf.committed = NFS_INVALID_STABLE_HOW;	/* not set yet */
 317	INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 318	dreq->mirror_count = 1;
 319	spin_lock_init(&dreq->lock);
 320
 321	return dreq;
 322}
 323
 324static void nfs_direct_req_free(struct kref *kref)
 325{
 326	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 327
 328	nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 329	if (dreq->l_ctx != NULL)
 330		nfs_put_lock_context(dreq->l_ctx);
 331	if (dreq->ctx != NULL)
 332		put_nfs_open_context(dreq->ctx);
 333	kmem_cache_free(nfs_direct_cachep, dreq);
 334}
 335
 336static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 337{
 338	kref_put(&dreq->kref, nfs_direct_req_free);
 339}
 340
 341ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 342{
 343	return dreq->bytes_left;
 
 344}
 345EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 346
 347/*
 348 * Collects and returns the final error value/byte-count.
 349 */
 350static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 351{
 352	ssize_t result = -EIOCBQUEUED;
 353
 354	/* Async requests don't wait here */
 355	if (dreq->iocb)
 356		goto out;
 357
 358	result = wait_for_completion_killable(&dreq->completion);
 359
 360	if (!result) {
 361		result = dreq->count;
 362		WARN_ON_ONCE(dreq->count < 0);
 363	}
 364	if (!result)
 365		result = dreq->error;
 366
 367out:
 368	return (ssize_t) result;
 369}
 370
 371/*
 372 * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 373 * the iocb is still valid here if this is a synchronous request.
 374 */
 375static void nfs_direct_complete(struct nfs_direct_req *dreq)
 376{
 377	struct inode *inode = dreq->inode;
 378
 379	inode_dio_end(inode);
 380
 381	if (dreq->iocb) {
 382		long res = (long) dreq->error;
 383		if (dreq->count != 0) {
 384			res = (long) dreq->count;
 385			WARN_ON_ONCE(dreq->count < 0);
 386		}
 387		dreq->iocb->ki_complete(dreq->iocb, res, 0);
 388	}
 389
 390	complete(&dreq->completion);
 391
 392	nfs_direct_req_release(dreq);
 393}
 394
 395static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 396{
 397	unsigned long bytes = 0;
 398	struct nfs_direct_req *dreq = hdr->dreq;
 399
 400	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 401		goto out_put;
 402
 403	spin_lock(&dreq->lock);
 404	if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 405		dreq->error = hdr->error;
 406	else
 407		nfs_direct_good_bytes(dreq, hdr);
 408
 
 409	spin_unlock(&dreq->lock);
 410
 411	while (!list_empty(&hdr->pages)) {
 412		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 413		struct page *page = req->wb_page;
 414
 415		if (!PageCompound(page) && bytes < hdr->good_bytes)
 
 416			set_page_dirty(page);
 417		bytes += req->wb_bytes;
 418		nfs_list_remove_request(req);
 419		nfs_release_request(req);
 420	}
 421out_put:
 422	if (put_dreq(dreq))
 423		nfs_direct_complete(dreq);
 424	hdr->release(hdr);
 425}
 426
 427static void nfs_read_sync_pgio_error(struct list_head *head)
 428{
 429	struct nfs_page *req;
 430
 431	while (!list_empty(head)) {
 432		req = nfs_list_entry(head->next);
 433		nfs_list_remove_request(req);
 434		nfs_release_request(req);
 435	}
 436}
 437
 438static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 439{
 440	get_dreq(hdr->dreq);
 441}
 442
 443static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 444	.error_cleanup = nfs_read_sync_pgio_error,
 445	.init_hdr = nfs_direct_pgio_init,
 446	.completion = nfs_direct_read_completion,
 447};
 448
 449/*
 450 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 451 * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 452 * bail and stop sending more reads.  Read length accounting is
 453 * handled automatically by nfs_direct_read_result().  Otherwise, if
 454 * no requests have been sent, just return an error.
 455 */
 456
 457static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 458					      struct iov_iter *iter,
 459					      loff_t pos)
 460{
 461	struct nfs_pageio_descriptor desc;
 462	struct inode *inode = dreq->inode;
 463	ssize_t result = -EINVAL;
 464	size_t requested_bytes = 0;
 465	size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 466
 467	nfs_pageio_init_read(&desc, dreq->inode, false,
 468			     &nfs_direct_read_completion_ops);
 469	get_dreq(dreq);
 470	desc.pg_dreq = dreq;
 471	inode_dio_begin(inode);
 472
 473	while (iov_iter_count(iter)) {
 474		struct page **pagevec;
 475		size_t bytes;
 476		size_t pgbase;
 477		unsigned npages, i;
 478
 479		result = iov_iter_get_pages_alloc(iter, &pagevec, 
 480						  rsize, &pgbase);
 481		if (result < 0)
 482			break;
 483	
 484		bytes = result;
 485		iov_iter_advance(iter, bytes);
 486		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 487		for (i = 0; i < npages; i++) {
 488			struct nfs_page *req;
 489			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 490			/* XXX do we need to do the eof zeroing found in async_filler? */
 491			req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 492						 pgbase, req_len);
 493			if (IS_ERR(req)) {
 494				result = PTR_ERR(req);
 495				break;
 496			}
 497			req->wb_index = pos >> PAGE_SHIFT;
 498			req->wb_offset = pos & ~PAGE_MASK;
 499			if (!nfs_pageio_add_request(&desc, req)) {
 500				result = desc.pg_error;
 501				nfs_release_request(req);
 502				break;
 503			}
 504			pgbase = 0;
 505			bytes -= req_len;
 506			requested_bytes += req_len;
 507			pos += req_len;
 508			dreq->bytes_left -= req_len;
 509		}
 510		nfs_direct_release_pages(pagevec, npages);
 511		kvfree(pagevec);
 512		if (result < 0)
 513			break;
 514	}
 515
 516	nfs_pageio_complete(&desc);
 517
 518	/*
 519	 * If no bytes were started, return the error, and let the
 520	 * generic layer handle the completion.
 521	 */
 522	if (requested_bytes == 0) {
 523		inode_dio_end(inode);
 524		nfs_direct_req_release(dreq);
 525		return result < 0 ? result : -EIO;
 526	}
 527
 528	if (put_dreq(dreq))
 529		nfs_direct_complete(dreq);
 530	return requested_bytes;
 531}
 532
 533/**
 534 * nfs_file_direct_read - file direct read operation for NFS files
 535 * @iocb: target I/O control block
 536 * @iter: vector of user buffers into which to read data
 
 537 *
 538 * We use this function for direct reads instead of calling
 539 * generic_file_aio_read() in order to avoid gfar's check to see if
 540 * the request starts before the end of the file.  For that check
 541 * to work, we must generate a GETATTR before each direct read, and
 542 * even then there is a window between the GETATTR and the subsequent
 543 * READ where the file size could change.  Our preference is simply
 544 * to do all reads the application wants, and the server will take
 545 * care of managing the end of file boundary.
 546 *
 547 * This function also eliminates unnecessarily updating the file's
 548 * atime locally, as the NFS server sets the file's atime, and this
 549 * client must read the updated atime from the server back into its
 550 * cache.
 551 */
 552ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
 
 553{
 554	struct file *file = iocb->ki_filp;
 555	struct address_space *mapping = file->f_mapping;
 556	struct inode *inode = mapping->host;
 557	struct nfs_direct_req *dreq;
 558	struct nfs_lock_context *l_ctx;
 559	ssize_t result = -EINVAL, requested;
 560	size_t count = iov_iter_count(iter);
 561	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 562
 563	dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 564		file, count, (long long) iocb->ki_pos);
 565
 566	result = 0;
 567	if (!count)
 568		goto out;
 569
 570	task_io_account_read(count);
 571
 572	result = -ENOMEM;
 573	dreq = nfs_direct_req_alloc();
 574	if (dreq == NULL)
 575		goto out;
 576
 577	dreq->inode = inode;
 578	dreq->bytes_left = dreq->max_count = count;
 579	dreq->io_start = iocb->ki_pos;
 580	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 581	l_ctx = nfs_get_lock_context(dreq->ctx);
 582	if (IS_ERR(l_ctx)) {
 583		result = PTR_ERR(l_ctx);
 
 584		goto out_release;
 585	}
 586	dreq->l_ctx = l_ctx;
 587	if (!is_sync_kiocb(iocb))
 588		dreq->iocb = iocb;
 589
 590	nfs_start_io_direct(inode);
 
 
 
 
 591
 592	NFS_I(inode)->read_io += count;
 593	requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 594
 595	nfs_end_io_direct(inode);
 
 596
 597	if (requested > 0) {
 598		result = nfs_direct_wait(dreq);
 599		if (result > 0) {
 600			requested -= result;
 601			iocb->ki_pos += result;
 602		}
 603		iov_iter_revert(iter, requested);
 604	} else {
 605		result = requested;
 606	}
 607
 608out_release:
 609	nfs_direct_req_release(dreq);
 610out:
 611	return result;
 612}
 613
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 614static void
 615nfs_direct_write_scan_commit_list(struct inode *inode,
 616				  struct list_head *list,
 617				  struct nfs_commit_info *cinfo)
 618{
 619	mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 620#ifdef CONFIG_NFS_V4_1
 621	if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 622		NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 623#endif
 624	nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 625	mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 626}
 627
 628static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 629{
 630	struct nfs_pageio_descriptor desc;
 631	struct nfs_page *req, *tmp;
 632	LIST_HEAD(reqs);
 633	struct nfs_commit_info cinfo;
 634	LIST_HEAD(failed);
 635	int i;
 636
 637	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 638	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 639
 640	dreq->count = 0;
 641	dreq->verf.committed = NFS_INVALID_STABLE_HOW;
 642	nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 643	for (i = 0; i < dreq->mirror_count; i++)
 644		dreq->mirrors[i].count = 0;
 645	get_dreq(dreq);
 646
 647	nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 648			      &nfs_direct_write_completion_ops);
 649	desc.pg_dreq = dreq;
 650
 651	req = nfs_list_entry(reqs.next);
 652	nfs_direct_setup_mirroring(dreq, &desc, req);
 653	if (desc.pg_error < 0) {
 654		list_splice_init(&reqs, &failed);
 655		goto out_failed;
 656	}
 657
 658	list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 659		if (!nfs_pageio_add_request(&desc, req)) {
 660			nfs_list_remove_request(req);
 661			nfs_list_add_request(req, &failed);
 662			spin_lock(&cinfo.inode->i_lock);
 663			dreq->flags = 0;
 664			if (desc.pg_error < 0)
 
 
 665				dreq->error = desc.pg_error;
 666			else
 667				dreq->error = -EIO;
 668			spin_unlock(&cinfo.inode->i_lock);
 
 669		}
 670		nfs_release_request(req);
 671	}
 672	nfs_pageio_complete(&desc);
 673
 674out_failed:
 675	while (!list_empty(&failed)) {
 676		req = nfs_list_entry(failed.next);
 677		nfs_list_remove_request(req);
 678		nfs_unlock_and_release_request(req);
 
 
 
 
 
 
 
 
 679	}
 680
 681	if (put_dreq(dreq))
 682		nfs_direct_write_complete(dreq);
 683}
 684
 685static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 686{
 
 687	struct nfs_direct_req *dreq = data->dreq;
 688	struct nfs_commit_info cinfo;
 689	struct nfs_page *req;
 690	int status = data->task.tk_status;
 691
 
 
 
 
 
 
 
 
 
 
 692	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 693	if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
 694		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 695
 696	while (!list_empty(&data->pages)) {
 697		req = nfs_list_entry(data->pages.next);
 698		nfs_list_remove_request(req);
 699		if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 700			/* Note the rewrite will go through mds */
 
 
 
 
 
 
 
 
 
 
 701			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 702		} else
 703			nfs_release_request(req);
 704		nfs_unlock_and_release_request(req);
 705	}
 706
 707	if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 708		nfs_direct_write_complete(dreq);
 709}
 710
 711static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 712		struct nfs_page *req)
 713{
 714	struct nfs_direct_req *dreq = cinfo->dreq;
 715
 
 
 716	spin_lock(&dreq->lock);
 717	dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 
 718	spin_unlock(&dreq->lock);
 719	nfs_mark_request_commit(req, NULL, cinfo, 0);
 720}
 721
 722static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 723	.completion = nfs_direct_commit_complete,
 724	.resched_write = nfs_direct_resched_write,
 725};
 726
 727static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 728{
 729	int res;
 730	struct nfs_commit_info cinfo;
 731	LIST_HEAD(mds_list);
 732
 733	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 734	nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 735	res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 736	if (res < 0) /* res == -ENOMEM */
 737		nfs_direct_write_reschedule(dreq);
 738}
 739
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 740static void nfs_direct_write_schedule_work(struct work_struct *work)
 741{
 742	struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 743	int flags = dreq->flags;
 744
 745	dreq->flags = 0;
 746	switch (flags) {
 747		case NFS_ODIRECT_DO_COMMIT:
 748			nfs_direct_commit_schedule(dreq);
 749			break;
 750		case NFS_ODIRECT_RESCHED_WRITES:
 751			nfs_direct_write_reschedule(dreq);
 752			break;
 753		default:
 
 754			nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 755			nfs_direct_complete(dreq);
 756	}
 757}
 758
 759static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 760{
 761	schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 
 762}
 763
 764static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 765{
 766	struct nfs_direct_req *dreq = hdr->dreq;
 767	struct nfs_commit_info cinfo;
 768	bool request_commit = false;
 769	struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 
 770
 771	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 772		goto out_put;
 773
 774	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 775
 776	spin_lock(&dreq->lock);
 
 
 
 
 777
 778	if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 779		dreq->error = hdr->error;
 780	if (dreq->error == 0) {
 781		nfs_direct_good_bytes(dreq, hdr);
 782		if (nfs_write_need_commit(hdr)) {
 783			if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 784				request_commit = true;
 785			else if (dreq->flags == 0) {
 786				nfs_direct_set_hdr_verf(dreq, hdr);
 787				request_commit = true;
 788				dreq->flags = NFS_ODIRECT_DO_COMMIT;
 789			} else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 790				request_commit = true;
 791				if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 792					dreq->flags =
 793						NFS_ODIRECT_RESCHED_WRITES;
 794			}
 795		}
 796	}
 797	spin_unlock(&dreq->lock);
 798
 799	while (!list_empty(&hdr->pages)) {
 800
 801		req = nfs_list_entry(hdr->pages.next);
 802		nfs_list_remove_request(req);
 803		if (request_commit) {
 804			kref_get(&req->wb_kref);
 
 
 805			nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 806				hdr->ds_commit_idx);
 
 
 
 807		}
 808		nfs_unlock_and_release_request(req);
 809	}
 810
 811out_put:
 812	if (put_dreq(dreq))
 813		nfs_direct_write_complete(dreq);
 814	hdr->release(hdr);
 815}
 816
 817static void nfs_write_sync_pgio_error(struct list_head *head)
 818{
 819	struct nfs_page *req;
 820
 821	while (!list_empty(head)) {
 822		req = nfs_list_entry(head->next);
 823		nfs_list_remove_request(req);
 824		nfs_unlock_and_release_request(req);
 825	}
 826}
 827
 828static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 829{
 830	struct nfs_direct_req *dreq = hdr->dreq;
 
 
 
 
 831
 
 832	spin_lock(&dreq->lock);
 833	if (dreq->error == 0) {
 834		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 835		/* fake unstable write to let common nfs resend pages */
 836		hdr->verf.committed = NFS_UNSTABLE;
 837		hdr->good_bytes = hdr->args.count;
 838	}
 839	spin_unlock(&dreq->lock);
 
 
 
 
 
 
 840}
 841
 842static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 843	.error_cleanup = nfs_write_sync_pgio_error,
 844	.init_hdr = nfs_direct_pgio_init,
 845	.completion = nfs_direct_write_completion,
 846	.reschedule_io = nfs_direct_write_reschedule_io,
 847};
 848
 849
 850/*
 851 * NB: Return the value of the first error return code.  Subsequent
 852 *     errors after the first one are ignored.
 853 */
 854/*
 855 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 856 * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 857 * bail and stop sending more writes.  Write length accounting is
 858 * handled automatically by nfs_direct_write_result().  Otherwise, if
 859 * no requests have been sent, just return an error.
 860 */
 861static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 862					       struct iov_iter *iter,
 863					       loff_t pos)
 864{
 865	struct nfs_pageio_descriptor desc;
 866	struct inode *inode = dreq->inode;
 
 867	ssize_t result = 0;
 868	size_t requested_bytes = 0;
 869	size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 
 
 
 870
 871	nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 872			      &nfs_direct_write_completion_ops);
 873	desc.pg_dreq = dreq;
 874	get_dreq(dreq);
 875	inode_dio_begin(inode);
 876
 877	NFS_I(inode)->write_io += iov_iter_count(iter);
 878	while (iov_iter_count(iter)) {
 879		struct page **pagevec;
 880		size_t bytes;
 881		size_t pgbase;
 882		unsigned npages, i;
 883
 884		result = iov_iter_get_pages_alloc(iter, &pagevec, 
 885						  wsize, &pgbase);
 886		if (result < 0)
 887			break;
 888
 889		bytes = result;
 890		iov_iter_advance(iter, bytes);
 891		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 892		for (i = 0; i < npages; i++) {
 893			struct nfs_page *req;
 894			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 895
 896			req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 897						 pgbase, req_len);
 898			if (IS_ERR(req)) {
 899				result = PTR_ERR(req);
 900				break;
 901			}
 902
 903			nfs_direct_setup_mirroring(dreq, &desc, req);
 904			if (desc.pg_error < 0) {
 905				nfs_free_request(req);
 906				result = desc.pg_error;
 907				break;
 908			}
 909
 
 
 
 
 
 
 
 
 
 
 910			nfs_lock_request(req);
 911			req->wb_index = pos >> PAGE_SHIFT;
 912			req->wb_offset = pos & ~PAGE_MASK;
 913			if (!nfs_pageio_add_request(&desc, req)) {
 
 
 914				result = desc.pg_error;
 915				nfs_unlock_and_release_request(req);
 916				break;
 917			}
 918			pgbase = 0;
 919			bytes -= req_len;
 920			requested_bytes += req_len;
 921			pos += req_len;
 922			dreq->bytes_left -= req_len;
 
 
 
 
 
 923		}
 924		nfs_direct_release_pages(pagevec, npages);
 925		kvfree(pagevec);
 926		if (result < 0)
 927			break;
 928	}
 929	nfs_pageio_complete(&desc);
 930
 931	/*
 932	 * If no bytes were started, return the error, and let the
 933	 * generic layer handle the completion.
 934	 */
 935	if (requested_bytes == 0) {
 936		inode_dio_end(inode);
 937		nfs_direct_req_release(dreq);
 938		return result < 0 ? result : -EIO;
 939	}
 940
 941	if (put_dreq(dreq))
 942		nfs_direct_write_complete(dreq);
 943	return requested_bytes;
 944}
 945
 946/**
 947 * nfs_file_direct_write - file direct write operation for NFS files
 948 * @iocb: target I/O control block
 949 * @iter: vector of user buffers from which to write data
 
 950 *
 951 * We use this function for direct writes instead of calling
 952 * generic_file_aio_write() in order to avoid taking the inode
 953 * semaphore and updating the i_size.  The NFS server will set
 954 * the new i_size and this client must read the updated size
 955 * back into its cache.  We let the server do generic write
 956 * parameter checking and report problems.
 957 *
 958 * We eliminate local atime updates, see direct read above.
 959 *
 960 * We avoid unnecessary page cache invalidations for normal cached
 961 * readers of this file.
 962 *
 963 * Note that O_APPEND is not supported for NFS direct writes, as there
 964 * is no atomic O_APPEND write facility in the NFS protocol.
 965 */
 966ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 
 967{
 968	ssize_t result = -EINVAL, requested;
 969	size_t count;
 970	struct file *file = iocb->ki_filp;
 971	struct address_space *mapping = file->f_mapping;
 972	struct inode *inode = mapping->host;
 973	struct nfs_direct_req *dreq;
 974	struct nfs_lock_context *l_ctx;
 975	loff_t pos, end;
 976
 977	dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 978		file, iov_iter_count(iter), (long long) iocb->ki_pos);
 979
 980	result = generic_write_checks(iocb, iter);
 
 
 
 
 981	if (result <= 0)
 982		return result;
 983	count = result;
 984	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 985
 986	pos = iocb->ki_pos;
 987	end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 988
 989	task_io_account_write(count);
 990
 991	result = -ENOMEM;
 992	dreq = nfs_direct_req_alloc();
 993	if (!dreq)
 994		goto out;
 995
 996	dreq->inode = inode;
 997	dreq->bytes_left = dreq->max_count = count;
 998	dreq->io_start = pos;
 999	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1000	l_ctx = nfs_get_lock_context(dreq->ctx);
1001	if (IS_ERR(l_ctx)) {
1002		result = PTR_ERR(l_ctx);
 
1003		goto out_release;
1004	}
1005	dreq->l_ctx = l_ctx;
1006	if (!is_sync_kiocb(iocb))
1007		dreq->iocb = iocb;
 
1008
1009	nfs_start_io_direct(inode);
 
 
 
 
1010
1011	requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
 
1012
1013	if (mapping->nrpages) {
1014		invalidate_inode_pages2_range(mapping,
1015					      pos >> PAGE_SHIFT, end);
1016	}
1017
1018	nfs_end_io_direct(inode);
 
1019
1020	if (requested > 0) {
1021		result = nfs_direct_wait(dreq);
1022		if (result > 0) {
1023			requested -= result;
1024			iocb->ki_pos = pos + result;
1025			/* XXX: should check the generic_write_sync retval */
1026			generic_write_sync(iocb, result);
1027		}
1028		iov_iter_revert(iter, requested);
1029	} else {
1030		result = requested;
1031	}
 
1032out_release:
1033	nfs_direct_req_release(dreq);
1034out:
1035	return result;
1036}
1037
1038/**
1039 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1040 *
1041 */
1042int __init nfs_init_directcache(void)
1043{
1044	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1045						sizeof(struct nfs_direct_req),
1046						0, (SLAB_RECLAIM_ACCOUNT|
1047							SLAB_MEM_SPREAD),
1048						NULL);
1049	if (nfs_direct_cachep == NULL)
1050		return -ENOMEM;
1051
1052	return 0;
1053}
1054
1055/**
1056 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1057 *
1058 */
1059void nfs_destroy_directcache(void)
1060{
1061	kmem_cache_destroy(nfs_direct_cachep);
1062}
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * linux/fs/nfs/direct.c
   4 *
   5 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   6 *
   7 * High-performance uncached I/O for the Linux NFS client
   8 *
   9 * There are important applications whose performance or correctness
  10 * depends on uncached access to file data.  Database clusters
  11 * (multiple copies of the same instance running on separate hosts)
  12 * implement their own cache coherency protocol that subsumes file
  13 * system cache protocols.  Applications that process datasets
  14 * considerably larger than the client's memory do not always benefit
  15 * from a local cache.  A streaming video server, for instance, has no
  16 * need to cache the contents of a file.
  17 *
  18 * When an application requests uncached I/O, all read and write requests
  19 * are made directly to the server; data stored or fetched via these
  20 * requests is not cached in the Linux page cache.  The client does not
  21 * correct unaligned requests from applications.  All requested bytes are
  22 * held on permanent storage before a direct write system call returns to
  23 * an application.
  24 *
  25 * Solaris implements an uncached I/O facility called directio() that
  26 * is used for backups and sequential I/O to very large files.  Solaris
  27 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  28 * an undocumented mount option.
  29 *
  30 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  31 * help from Andrew Morton.
  32 *
  33 * 18 Dec 2001	Initial implementation for 2.4  --cel
  34 * 08 Jul 2002	Version for 2.4.19, with bug fixes --trondmy
  35 * 08 Jun 2003	Port to 2.5 APIs  --cel
  36 * 31 Mar 2004	Handle direct I/O without VFS support  --cel
  37 * 15 Sep 2004	Parallel async reads  --cel
  38 * 04 May 2005	support O_DIRECT with aio  --cel
  39 *
  40 */
  41
  42#include <linux/errno.h>
  43#include <linux/sched.h>
  44#include <linux/kernel.h>
  45#include <linux/file.h>
  46#include <linux/pagemap.h>
  47#include <linux/kref.h>
  48#include <linux/slab.h>
  49#include <linux/task_io_accounting_ops.h>
  50#include <linux/module.h>
  51
  52#include <linux/nfs_fs.h>
  53#include <linux/nfs_page.h>
  54#include <linux/sunrpc/clnt.h>
  55
  56#include <linux/uaccess.h>
  57#include <linux/atomic.h>
  58
  59#include "internal.h"
  60#include "iostat.h"
  61#include "pnfs.h"
  62#include "fscache.h"
  63#include "nfstrace.h"
  64
  65#define NFSDBG_FACILITY		NFSDBG_VFS
  66
  67static struct kmem_cache *nfs_direct_cachep;
  68
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  69static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  70static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  71static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
  72static void nfs_direct_write_schedule_work(struct work_struct *work);
  73
  74static inline void get_dreq(struct nfs_direct_req *dreq)
  75{
  76	atomic_inc(&dreq->io_count);
  77}
  78
  79static inline int put_dreq(struct nfs_direct_req *dreq)
  80{
  81	return atomic_dec_and_test(&dreq->io_count);
  82}
  83
  84static void
  85nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
  86			    const struct nfs_pgio_header *hdr,
  87			    ssize_t dreq_len)
  88{
  89	if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
  90	      test_bit(NFS_IOHDR_EOF, &hdr->flags)))
  91		return;
  92	if (dreq->max_count >= dreq_len) {
  93		dreq->max_count = dreq_len;
  94		if (dreq->count > dreq_len)
  95			dreq->count = dreq_len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  96	}
 
  97
  98	if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && !dreq->error)
  99		dreq->error = hdr->error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 100}
 101
 102static void
 103nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 104		       const struct nfs_pgio_header *hdr)
 
 
 
 
 
 
 
 105{
 106	loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 107	ssize_t dreq_len = 0;
 108
 109	if (hdr_end > dreq->io_start)
 110		dreq_len = hdr_end - dreq->io_start;
 
 
 
 111
 112	nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 
 
 
 
 113
 114	if (dreq_len > dreq->max_count)
 115		dreq_len = dreq->max_count;
 
 
 
 
 
 
 
 
 
 
 
 116
 117	if (dreq->count < dreq_len)
 118		dreq->count = dreq_len;
 
 
 
 
 119}
 120
 121static void nfs_direct_truncate_request(struct nfs_direct_req *dreq,
 122					struct nfs_page *req)
 
 
 
 
 
 
 
 
 123{
 124	loff_t offs = req_offset(req);
 125	size_t req_start = (size_t)(offs - dreq->io_start);
 126
 127	if (req_start < dreq->max_count)
 128		dreq->max_count = req_start;
 129	if (req_start < dreq->count)
 130		dreq->count = req_start;
 
 
 
 
 131}
 132
 133/**
 134 * nfs_swap_rw - NFS address space operation for swap I/O
 135 * @iocb: target I/O control block
 136 * @iter: I/O buffer
 137 *
 138 * Perform IO to the swap-file.  This is much like direct IO.
 
 
 
 139 */
 140int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
 141{
 142	ssize_t ret;
 
 
 
 
 143
 144	VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 145
 146	if (iov_iter_rw(iter) == READ)
 147		ret = nfs_file_direct_read(iocb, iter, true);
 148	else
 149		ret = nfs_file_direct_write(iocb, iter, true);
 150	if (ret < 0)
 151		return ret;
 152	return 0;
 153}
 154
 155static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 156{
 157	unsigned int i;
 158	for (i = 0; i < npages; i++)
 159		put_page(pages[i]);
 160}
 161
 162void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 163			      struct nfs_direct_req *dreq)
 164{
 165	cinfo->inode = dreq->inode;
 166	cinfo->mds = &dreq->mds_cinfo;
 167	cinfo->ds = &dreq->ds_cinfo;
 168	cinfo->dreq = dreq;
 169	cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 170}
 171
 
 
 
 
 
 
 
 
 
 
 
 
 172static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 173{
 174	struct nfs_direct_req *dreq;
 175
 176	dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 177	if (!dreq)
 178		return NULL;
 179
 180	kref_init(&dreq->kref);
 181	kref_get(&dreq->kref);
 182	init_completion(&dreq->completion);
 183	INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 184	pnfs_init_ds_commit_info(&dreq->ds_cinfo);
 185	INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 
 186	spin_lock_init(&dreq->lock);
 187
 188	return dreq;
 189}
 190
 191static void nfs_direct_req_free(struct kref *kref)
 192{
 193	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 194
 195	pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
 196	if (dreq->l_ctx != NULL)
 197		nfs_put_lock_context(dreq->l_ctx);
 198	if (dreq->ctx != NULL)
 199		put_nfs_open_context(dreq->ctx);
 200	kmem_cache_free(nfs_direct_cachep, dreq);
 201}
 202
 203static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 204{
 205	kref_put(&dreq->kref, nfs_direct_req_free);
 206}
 207
 208ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq, loff_t offset)
 209{
 210	loff_t start = offset - dreq->io_start;
 211	return dreq->max_count - start;
 212}
 213EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 214
 215/*
 216 * Collects and returns the final error value/byte-count.
 217 */
 218static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 219{
 220	ssize_t result = -EIOCBQUEUED;
 221
 222	/* Async requests don't wait here */
 223	if (dreq->iocb)
 224		goto out;
 225
 226	result = wait_for_completion_killable(&dreq->completion);
 227
 228	if (!result) {
 229		result = dreq->count;
 230		WARN_ON_ONCE(dreq->count < 0);
 231	}
 232	if (!result)
 233		result = dreq->error;
 234
 235out:
 236	return (ssize_t) result;
 237}
 238
 239/*
 240 * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 241 * the iocb is still valid here if this is a synchronous request.
 242 */
 243static void nfs_direct_complete(struct nfs_direct_req *dreq)
 244{
 245	struct inode *inode = dreq->inode;
 246
 247	inode_dio_end(inode);
 248
 249	if (dreq->iocb) {
 250		long res = (long) dreq->error;
 251		if (dreq->count != 0) {
 252			res = (long) dreq->count;
 253			WARN_ON_ONCE(dreq->count < 0);
 254		}
 255		dreq->iocb->ki_complete(dreq->iocb, res);
 256	}
 257
 258	complete(&dreq->completion);
 259
 260	nfs_direct_req_release(dreq);
 261}
 262
 263static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 264{
 265	unsigned long bytes = 0;
 266	struct nfs_direct_req *dreq = hdr->dreq;
 267
 
 
 
 268	spin_lock(&dreq->lock);
 269	if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 270		spin_unlock(&dreq->lock);
 271		goto out_put;
 272	}
 273
 274	nfs_direct_count_bytes(dreq, hdr);
 275	spin_unlock(&dreq->lock);
 276
 277	while (!list_empty(&hdr->pages)) {
 278		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 279		struct page *page = req->wb_page;
 280
 281		if (!PageCompound(page) && bytes < hdr->good_bytes &&
 282		    (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 283			set_page_dirty(page);
 284		bytes += req->wb_bytes;
 285		nfs_list_remove_request(req);
 286		nfs_release_request(req);
 287	}
 288out_put:
 289	if (put_dreq(dreq))
 290		nfs_direct_complete(dreq);
 291	hdr->release(hdr);
 292}
 293
 294static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 295{
 296	struct nfs_page *req;
 297
 298	while (!list_empty(head)) {
 299		req = nfs_list_entry(head->next);
 300		nfs_list_remove_request(req);
 301		nfs_release_request(req);
 302	}
 303}
 304
 305static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 306{
 307	get_dreq(hdr->dreq);
 308}
 309
 310static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 311	.error_cleanup = nfs_read_sync_pgio_error,
 312	.init_hdr = nfs_direct_pgio_init,
 313	.completion = nfs_direct_read_completion,
 314};
 315
 316/*
 317 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 318 * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 319 * bail and stop sending more reads.  Read length accounting is
 320 * handled automatically by nfs_direct_read_result().  Otherwise, if
 321 * no requests have been sent, just return an error.
 322 */
 323
 324static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 325					      struct iov_iter *iter,
 326					      loff_t pos)
 327{
 328	struct nfs_pageio_descriptor desc;
 329	struct inode *inode = dreq->inode;
 330	ssize_t result = -EINVAL;
 331	size_t requested_bytes = 0;
 332	size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 333
 334	nfs_pageio_init_read(&desc, dreq->inode, false,
 335			     &nfs_direct_read_completion_ops);
 336	get_dreq(dreq);
 337	desc.pg_dreq = dreq;
 338	inode_dio_begin(inode);
 339
 340	while (iov_iter_count(iter)) {
 341		struct page **pagevec;
 342		size_t bytes;
 343		size_t pgbase;
 344		unsigned npages, i;
 345
 346		result = iov_iter_get_pages_alloc2(iter, &pagevec,
 347						  rsize, &pgbase);
 348		if (result < 0)
 349			break;
 350	
 351		bytes = result;
 
 352		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 353		for (i = 0; i < npages; i++) {
 354			struct nfs_page *req;
 355			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 356			/* XXX do we need to do the eof zeroing found in async_filler? */
 357			req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 358							pgbase, pos, req_len);
 359			if (IS_ERR(req)) {
 360				result = PTR_ERR(req);
 361				break;
 362			}
 
 
 363			if (!nfs_pageio_add_request(&desc, req)) {
 364				result = desc.pg_error;
 365				nfs_release_request(req);
 366				break;
 367			}
 368			pgbase = 0;
 369			bytes -= req_len;
 370			requested_bytes += req_len;
 371			pos += req_len;
 
 372		}
 373		nfs_direct_release_pages(pagevec, npages);
 374		kvfree(pagevec);
 375		if (result < 0)
 376			break;
 377	}
 378
 379	nfs_pageio_complete(&desc);
 380
 381	/*
 382	 * If no bytes were started, return the error, and let the
 383	 * generic layer handle the completion.
 384	 */
 385	if (requested_bytes == 0) {
 386		inode_dio_end(inode);
 387		nfs_direct_req_release(dreq);
 388		return result < 0 ? result : -EIO;
 389	}
 390
 391	if (put_dreq(dreq))
 392		nfs_direct_complete(dreq);
 393	return requested_bytes;
 394}
 395
 396/**
 397 * nfs_file_direct_read - file direct read operation for NFS files
 398 * @iocb: target I/O control block
 399 * @iter: vector of user buffers into which to read data
 400 * @swap: flag indicating this is swap IO, not O_DIRECT IO
 401 *
 402 * We use this function for direct reads instead of calling
 403 * generic_file_aio_read() in order to avoid gfar's check to see if
 404 * the request starts before the end of the file.  For that check
 405 * to work, we must generate a GETATTR before each direct read, and
 406 * even then there is a window between the GETATTR and the subsequent
 407 * READ where the file size could change.  Our preference is simply
 408 * to do all reads the application wants, and the server will take
 409 * care of managing the end of file boundary.
 410 *
 411 * This function also eliminates unnecessarily updating the file's
 412 * atime locally, as the NFS server sets the file's atime, and this
 413 * client must read the updated atime from the server back into its
 414 * cache.
 415 */
 416ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 417			     bool swap)
 418{
 419	struct file *file = iocb->ki_filp;
 420	struct address_space *mapping = file->f_mapping;
 421	struct inode *inode = mapping->host;
 422	struct nfs_direct_req *dreq;
 423	struct nfs_lock_context *l_ctx;
 424	ssize_t result, requested;
 425	size_t count = iov_iter_count(iter);
 426	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 427
 428	dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 429		file, count, (long long) iocb->ki_pos);
 430
 431	result = 0;
 432	if (!count)
 433		goto out;
 434
 435	task_io_account_read(count);
 436
 437	result = -ENOMEM;
 438	dreq = nfs_direct_req_alloc();
 439	if (dreq == NULL)
 440		goto out;
 441
 442	dreq->inode = inode;
 443	dreq->max_count = count;
 444	dreq->io_start = iocb->ki_pos;
 445	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 446	l_ctx = nfs_get_lock_context(dreq->ctx);
 447	if (IS_ERR(l_ctx)) {
 448		result = PTR_ERR(l_ctx);
 449		nfs_direct_req_release(dreq);
 450		goto out_release;
 451	}
 452	dreq->l_ctx = l_ctx;
 453	if (!is_sync_kiocb(iocb))
 454		dreq->iocb = iocb;
 455
 456	if (user_backed_iter(iter))
 457		dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 458
 459	if (!swap)
 460		nfs_start_io_direct(inode);
 461
 462	NFS_I(inode)->read_io += count;
 463	requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 464
 465	if (!swap)
 466		nfs_end_io_direct(inode);
 467
 468	if (requested > 0) {
 469		result = nfs_direct_wait(dreq);
 470		if (result > 0) {
 471			requested -= result;
 472			iocb->ki_pos += result;
 473		}
 474		iov_iter_revert(iter, requested);
 475	} else {
 476		result = requested;
 477	}
 478
 479out_release:
 480	nfs_direct_req_release(dreq);
 481out:
 482	return result;
 483}
 484
 485static void nfs_direct_add_page_head(struct list_head *list,
 486				     struct nfs_page *req)
 487{
 488	struct nfs_page *head = req->wb_head;
 489
 490	if (!list_empty(&head->wb_list) || !nfs_lock_request(head))
 491		return;
 492	if (!list_empty(&head->wb_list)) {
 493		nfs_unlock_request(head);
 494		return;
 495	}
 496	list_add(&head->wb_list, list);
 497	kref_get(&head->wb_kref);
 498	kref_get(&head->wb_kref);
 499}
 500
 501static void nfs_direct_join_group(struct list_head *list,
 502				  struct nfs_commit_info *cinfo,
 503				  struct inode *inode)
 504{
 505	struct nfs_page *req, *subreq;
 506
 507	list_for_each_entry(req, list, wb_list) {
 508		if (req->wb_head != req) {
 509			nfs_direct_add_page_head(&req->wb_list, req);
 510			continue;
 511		}
 512		subreq = req->wb_this_page;
 513		if (subreq == req)
 514			continue;
 515		do {
 516			/*
 517			 * Remove subrequests from this list before freeing
 518			 * them in the call to nfs_join_page_group().
 519			 */
 520			if (!list_empty(&subreq->wb_list)) {
 521				nfs_list_remove_request(subreq);
 522				nfs_release_request(subreq);
 523			}
 524		} while ((subreq = subreq->wb_this_page) != req);
 525		nfs_join_page_group(req, cinfo, inode);
 526	}
 527}
 528
 529static void
 530nfs_direct_write_scan_commit_list(struct inode *inode,
 531				  struct list_head *list,
 532				  struct nfs_commit_info *cinfo)
 533{
 534	mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 535	pnfs_recover_commit_reqs(list, cinfo);
 
 
 
 536	nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 537	mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 538}
 539
 540static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 541{
 542	struct nfs_pageio_descriptor desc;
 543	struct nfs_page *req;
 544	LIST_HEAD(reqs);
 545	struct nfs_commit_info cinfo;
 
 
 546
 547	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 548	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 549
 550	nfs_direct_join_group(&reqs, &cinfo, dreq->inode);
 551
 552	nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 
 
 553	get_dreq(dreq);
 554
 555	nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 556			      &nfs_direct_write_completion_ops);
 557	desc.pg_dreq = dreq;
 558
 559	while (!list_empty(&reqs)) {
 560		req = nfs_list_entry(reqs.next);
 561		/* Bump the transmission count */
 562		req->wb_nio++;
 
 
 
 
 563		if (!nfs_pageio_add_request(&desc, req)) {
 564			spin_lock(&dreq->lock);
 565			if (dreq->error < 0) {
 566				desc.pg_error = dreq->error;
 567			} else if (desc.pg_error != -EAGAIN) {
 568				dreq->flags = 0;
 569				if (!desc.pg_error)
 570					desc.pg_error = -EIO;
 571				dreq->error = desc.pg_error;
 572			} else
 573				dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 574			spin_unlock(&dreq->lock);
 575			break;
 576		}
 577		nfs_release_request(req);
 578	}
 579	nfs_pageio_complete(&desc);
 580
 581	while (!list_empty(&reqs)) {
 582		req = nfs_list_entry(reqs.next);
 
 583		nfs_list_remove_request(req);
 584		nfs_unlock_and_release_request(req);
 585		if (desc.pg_error == -EAGAIN) {
 586			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 587		} else {
 588			spin_lock(&dreq->lock);
 589			nfs_direct_truncate_request(dreq, req);
 590			spin_unlock(&dreq->lock);
 591			nfs_release_request(req);
 592		}
 593	}
 594
 595	if (put_dreq(dreq))
 596		nfs_direct_write_complete(dreq);
 597}
 598
 599static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 600{
 601	const struct nfs_writeverf *verf = data->res.verf;
 602	struct nfs_direct_req *dreq = data->dreq;
 603	struct nfs_commit_info cinfo;
 604	struct nfs_page *req;
 605	int status = data->task.tk_status;
 606
 607	trace_nfs_direct_commit_complete(dreq);
 608
 609	if (status < 0) {
 610		/* Errors in commit are fatal */
 611		dreq->error = status;
 612		dreq->flags = NFS_ODIRECT_DONE;
 613	} else {
 614		status = dreq->error;
 615	}
 616
 617	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 
 
 618
 619	while (!list_empty(&data->pages)) {
 620		req = nfs_list_entry(data->pages.next);
 621		nfs_list_remove_request(req);
 622		if (status < 0) {
 623			spin_lock(&dreq->lock);
 624			nfs_direct_truncate_request(dreq, req);
 625			spin_unlock(&dreq->lock);
 626			nfs_release_request(req);
 627		} else if (!nfs_write_match_verf(verf, req)) {
 628			dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 629			/*
 630			 * Despite the reboot, the write was successful,
 631			 * so reset wb_nio.
 632			 */
 633			req->wb_nio = 0;
 634			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 635		} else
 636			nfs_release_request(req);
 637		nfs_unlock_and_release_request(req);
 638	}
 639
 640	if (nfs_commit_end(cinfo.mds))
 641		nfs_direct_write_complete(dreq);
 642}
 643
 644static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 645		struct nfs_page *req)
 646{
 647	struct nfs_direct_req *dreq = cinfo->dreq;
 648
 649	trace_nfs_direct_resched_write(dreq);
 650
 651	spin_lock(&dreq->lock);
 652	if (dreq->flags != NFS_ODIRECT_DONE)
 653		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 654	spin_unlock(&dreq->lock);
 655	nfs_mark_request_commit(req, NULL, cinfo, 0);
 656}
 657
 658static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 659	.completion = nfs_direct_commit_complete,
 660	.resched_write = nfs_direct_resched_write,
 661};
 662
 663static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 664{
 665	int res;
 666	struct nfs_commit_info cinfo;
 667	LIST_HEAD(mds_list);
 668
 669	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 670	nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 671	res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 672	if (res < 0) /* res == -ENOMEM */
 673		nfs_direct_write_reschedule(dreq);
 674}
 675
 676static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
 677{
 678	struct nfs_commit_info cinfo;
 679	struct nfs_page *req;
 680	LIST_HEAD(reqs);
 681
 682	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 683	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 684
 685	while (!list_empty(&reqs)) {
 686		req = nfs_list_entry(reqs.next);
 687		nfs_list_remove_request(req);
 688		nfs_direct_truncate_request(dreq, req);
 689		nfs_release_request(req);
 690		nfs_unlock_and_release_request(req);
 691	}
 692}
 693
 694static void nfs_direct_write_schedule_work(struct work_struct *work)
 695{
 696	struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 697	int flags = dreq->flags;
 698
 699	dreq->flags = 0;
 700	switch (flags) {
 701		case NFS_ODIRECT_DO_COMMIT:
 702			nfs_direct_commit_schedule(dreq);
 703			break;
 704		case NFS_ODIRECT_RESCHED_WRITES:
 705			nfs_direct_write_reschedule(dreq);
 706			break;
 707		default:
 708			nfs_direct_write_clear_reqs(dreq);
 709			nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 710			nfs_direct_complete(dreq);
 711	}
 712}
 713
 714static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 715{
 716	trace_nfs_direct_write_complete(dreq);
 717	queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 718}
 719
 720static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 721{
 722	struct nfs_direct_req *dreq = hdr->dreq;
 723	struct nfs_commit_info cinfo;
 
 724	struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 725	int flags = NFS_ODIRECT_DONE;
 726
 727	trace_nfs_direct_write_completion(dreq);
 
 728
 729	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 730
 731	spin_lock(&dreq->lock);
 732	if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 733		spin_unlock(&dreq->lock);
 734		goto out_put;
 735	}
 736
 737	nfs_direct_count_bytes(dreq, hdr);
 738	if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags) &&
 739	    !test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 740		if (!dreq->flags)
 741			dreq->flags = NFS_ODIRECT_DO_COMMIT;
 742		flags = dreq->flags;
 
 
 
 
 
 
 
 
 
 
 
 
 743	}
 744	spin_unlock(&dreq->lock);
 745
 746	while (!list_empty(&hdr->pages)) {
 747
 748		req = nfs_list_entry(hdr->pages.next);
 749		nfs_list_remove_request(req);
 750		if (flags == NFS_ODIRECT_DO_COMMIT) {
 751			kref_get(&req->wb_kref);
 752			memcpy(&req->wb_verf, &hdr->verf.verifier,
 753			       sizeof(req->wb_verf));
 754			nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 755				hdr->ds_commit_idx);
 756		} else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
 757			kref_get(&req->wb_kref);
 758			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 759		}
 760		nfs_unlock_and_release_request(req);
 761	}
 762
 763out_put:
 764	if (put_dreq(dreq))
 765		nfs_direct_write_complete(dreq);
 766	hdr->release(hdr);
 767}
 768
 769static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 770{
 771	struct nfs_page *req;
 772
 773	while (!list_empty(head)) {
 774		req = nfs_list_entry(head->next);
 775		nfs_list_remove_request(req);
 776		nfs_unlock_and_release_request(req);
 777	}
 778}
 779
 780static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 781{
 782	struct nfs_direct_req *dreq = hdr->dreq;
 783	struct nfs_page *req;
 784	struct nfs_commit_info cinfo;
 785
 786	trace_nfs_direct_write_reschedule_io(dreq);
 787
 788	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 789	spin_lock(&dreq->lock);
 790	if (dreq->error == 0)
 791		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 792	set_bit(NFS_IOHDR_REDO, &hdr->flags);
 
 
 
 793	spin_unlock(&dreq->lock);
 794	while (!list_empty(&hdr->pages)) {
 795		req = nfs_list_entry(hdr->pages.next);
 796		nfs_list_remove_request(req);
 797		nfs_unlock_request(req);
 798		nfs_mark_request_commit(req, NULL, &cinfo, 0);
 799	}
 800}
 801
 802static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 803	.error_cleanup = nfs_write_sync_pgio_error,
 804	.init_hdr = nfs_direct_pgio_init,
 805	.completion = nfs_direct_write_completion,
 806	.reschedule_io = nfs_direct_write_reschedule_io,
 807};
 808
 809
 810/*
 811 * NB: Return the value of the first error return code.  Subsequent
 812 *     errors after the first one are ignored.
 813 */
 814/*
 815 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 816 * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 817 * bail and stop sending more writes.  Write length accounting is
 818 * handled automatically by nfs_direct_write_result().  Otherwise, if
 819 * no requests have been sent, just return an error.
 820 */
 821static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 822					       struct iov_iter *iter,
 823					       loff_t pos, int ioflags)
 824{
 825	struct nfs_pageio_descriptor desc;
 826	struct inode *inode = dreq->inode;
 827	struct nfs_commit_info cinfo;
 828	ssize_t result = 0;
 829	size_t requested_bytes = 0;
 830	size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 831	bool defer = false;
 832
 833	trace_nfs_direct_write_schedule_iovec(dreq);
 834
 835	nfs_pageio_init_write(&desc, inode, ioflags, false,
 836			      &nfs_direct_write_completion_ops);
 837	desc.pg_dreq = dreq;
 838	get_dreq(dreq);
 839	inode_dio_begin(inode);
 840
 841	NFS_I(inode)->write_io += iov_iter_count(iter);
 842	while (iov_iter_count(iter)) {
 843		struct page **pagevec;
 844		size_t bytes;
 845		size_t pgbase;
 846		unsigned npages, i;
 847
 848		result = iov_iter_get_pages_alloc2(iter, &pagevec,
 849						  wsize, &pgbase);
 850		if (result < 0)
 851			break;
 852
 853		bytes = result;
 
 854		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 855		for (i = 0; i < npages; i++) {
 856			struct nfs_page *req;
 857			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 858
 859			req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 860							pgbase, pos, req_len);
 861			if (IS_ERR(req)) {
 862				result = PTR_ERR(req);
 863				break;
 864			}
 865
 
 866			if (desc.pg_error < 0) {
 867				nfs_free_request(req);
 868				result = desc.pg_error;
 869				break;
 870			}
 871
 872			pgbase = 0;
 873			bytes -= req_len;
 874			requested_bytes += req_len;
 875			pos += req_len;
 876
 877			if (defer) {
 878				nfs_mark_request_commit(req, NULL, &cinfo, 0);
 879				continue;
 880			}
 881
 882			nfs_lock_request(req);
 883			if (nfs_pageio_add_request(&desc, req))
 884				continue;
 885
 886			/* Exit on hard errors */
 887			if (desc.pg_error < 0 && desc.pg_error != -EAGAIN) {
 888				result = desc.pg_error;
 889				nfs_unlock_and_release_request(req);
 890				break;
 891			}
 892
 893			/* If the error is soft, defer remaining requests */
 894			nfs_init_cinfo_from_dreq(&cinfo, dreq);
 895			spin_lock(&dreq->lock);
 896			dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 897			spin_unlock(&dreq->lock);
 898			nfs_unlock_request(req);
 899			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 900			desc.pg_error = 0;
 901			defer = true;
 902		}
 903		nfs_direct_release_pages(pagevec, npages);
 904		kvfree(pagevec);
 905		if (result < 0)
 906			break;
 907	}
 908	nfs_pageio_complete(&desc);
 909
 910	/*
 911	 * If no bytes were started, return the error, and let the
 912	 * generic layer handle the completion.
 913	 */
 914	if (requested_bytes == 0) {
 915		inode_dio_end(inode);
 916		nfs_direct_req_release(dreq);
 917		return result < 0 ? result : -EIO;
 918	}
 919
 920	if (put_dreq(dreq))
 921		nfs_direct_write_complete(dreq);
 922	return requested_bytes;
 923}
 924
 925/**
 926 * nfs_file_direct_write - file direct write operation for NFS files
 927 * @iocb: target I/O control block
 928 * @iter: vector of user buffers from which to write data
 929 * @swap: flag indicating this is swap IO, not O_DIRECT IO
 930 *
 931 * We use this function for direct writes instead of calling
 932 * generic_file_aio_write() in order to avoid taking the inode
 933 * semaphore and updating the i_size.  The NFS server will set
 934 * the new i_size and this client must read the updated size
 935 * back into its cache.  We let the server do generic write
 936 * parameter checking and report problems.
 937 *
 938 * We eliminate local atime updates, see direct read above.
 939 *
 940 * We avoid unnecessary page cache invalidations for normal cached
 941 * readers of this file.
 942 *
 943 * Note that O_APPEND is not supported for NFS direct writes, as there
 944 * is no atomic O_APPEND write facility in the NFS protocol.
 945 */
 946ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 947			      bool swap)
 948{
 949	ssize_t result, requested;
 950	size_t count;
 951	struct file *file = iocb->ki_filp;
 952	struct address_space *mapping = file->f_mapping;
 953	struct inode *inode = mapping->host;
 954	struct nfs_direct_req *dreq;
 955	struct nfs_lock_context *l_ctx;
 956	loff_t pos, end;
 957
 958	dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 959		file, iov_iter_count(iter), (long long) iocb->ki_pos);
 960
 961	if (swap)
 962		/* bypass generic checks */
 963		result =  iov_iter_count(iter);
 964	else
 965		result = generic_write_checks(iocb, iter);
 966	if (result <= 0)
 967		return result;
 968	count = result;
 969	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 970
 971	pos = iocb->ki_pos;
 972	end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 973
 974	task_io_account_write(count);
 975
 976	result = -ENOMEM;
 977	dreq = nfs_direct_req_alloc();
 978	if (!dreq)
 979		goto out;
 980
 981	dreq->inode = inode;
 982	dreq->max_count = count;
 983	dreq->io_start = pos;
 984	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 985	l_ctx = nfs_get_lock_context(dreq->ctx);
 986	if (IS_ERR(l_ctx)) {
 987		result = PTR_ERR(l_ctx);
 988		nfs_direct_req_release(dreq);
 989		goto out_release;
 990	}
 991	dreq->l_ctx = l_ctx;
 992	if (!is_sync_kiocb(iocb))
 993		dreq->iocb = iocb;
 994	pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
 995
 996	if (swap) {
 997		requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
 998							    FLUSH_STABLE);
 999	} else {
1000		nfs_start_io_direct(inode);
1001
1002		requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1003							    FLUSH_COND_STABLE);
1004
1005		if (mapping->nrpages) {
1006			invalidate_inode_pages2_range(mapping,
1007						      pos >> PAGE_SHIFT, end);
1008		}
1009
1010		nfs_end_io_direct(inode);
1011	}
1012
1013	if (requested > 0) {
1014		result = nfs_direct_wait(dreq);
1015		if (result > 0) {
1016			requested -= result;
1017			iocb->ki_pos = pos + result;
1018			/* XXX: should check the generic_write_sync retval */
1019			generic_write_sync(iocb, result);
1020		}
1021		iov_iter_revert(iter, requested);
1022	} else {
1023		result = requested;
1024	}
1025	nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
1026out_release:
1027	nfs_direct_req_release(dreq);
1028out:
1029	return result;
1030}
1031
1032/**
1033 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1034 *
1035 */
1036int __init nfs_init_directcache(void)
1037{
1038	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1039						sizeof(struct nfs_direct_req),
1040						0, (SLAB_RECLAIM_ACCOUNT|
1041							SLAB_MEM_SPREAD),
1042						NULL);
1043	if (nfs_direct_cachep == NULL)
1044		return -ENOMEM;
1045
1046	return 0;
1047}
1048
1049/**
1050 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1051 *
1052 */
1053void nfs_destroy_directcache(void)
1054{
1055	kmem_cache_destroy(nfs_direct_cachep);
1056}