Linux Audio

Check our new training course

Loading...
v5.4
   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
  63#define NFSDBG_FACILITY		NFSDBG_VFS
  64
  65static struct kmem_cache *nfs_direct_cachep;
  66
 
 
 
 
 
 
 
  67struct nfs_direct_req {
  68	struct kref		kref;		/* release manager */
  69
  70	/* I/O parameters */
  71	struct nfs_open_context	*ctx;		/* file open context info */
  72	struct nfs_lock_context *l_ctx;		/* Lock context info */
  73	struct kiocb *		iocb;		/* controlling i/o request */
  74	struct inode *		inode;		/* target file of i/o */
  75
  76	/* completion state */
  77	atomic_t		io_count;	/* i/os we're waiting for */
  78	spinlock_t		lock;		/* protect completion state */
  79
 
 
 
  80	loff_t			io_start;	/* Start offset for I/O */
  81	ssize_t			count,		/* bytes actually processed */
  82				max_count,	/* max expected count */
  83				bytes_left,	/* bytes left to be sent */
  84				error;		/* any reported error */
  85	struct completion	completion;	/* wait for i/o completion */
  86
  87	/* commit state */
  88	struct nfs_mds_commit_info mds_cinfo;	/* Storage for cinfo */
  89	struct pnfs_ds_commit_info ds_cinfo;	/* Storage for cinfo */
  90	struct work_struct	work;
  91	int			flags;
  92	/* for write */
  93#define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
  94#define NFS_ODIRECT_RESCHED_WRITES	(2)	/* write verification failed */
  95	/* for read */
  96#define NFS_ODIRECT_SHOULD_DIRTY	(3)	/* dirty user-space page after read */
  97	struct nfs_writeverf	verf;		/* unstable write verifier */
  98};
  99
 100static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 101static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 102static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
 103static void nfs_direct_write_schedule_work(struct work_struct *work);
 104
 105static inline void get_dreq(struct nfs_direct_req *dreq)
 106{
 107	atomic_inc(&dreq->io_count);
 108}
 109
 110static inline int put_dreq(struct nfs_direct_req *dreq)
 111{
 112	return atomic_dec_and_test(&dreq->io_count);
 113}
 114
 115static void
 116nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
 117			    const struct nfs_pgio_header *hdr,
 118			    ssize_t dreq_len)
 119{
 120	if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
 121	      test_bit(NFS_IOHDR_EOF, &hdr->flags)))
 122		return;
 123	if (dreq->max_count >= dreq_len) {
 124		dreq->max_count = dreq_len;
 125		if (dreq->count > dreq_len)
 126			dreq->count = dreq_len;
 127
 128		if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 129			dreq->error = hdr->error;
 130		else /* Clear outstanding error if this is EOF */
 131			dreq->error = 0;
 132	}
 133}
 134
 135static void
 136nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 137		       const struct nfs_pgio_header *hdr)
 138{
 139	loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 140	ssize_t dreq_len = 0;
 141
 142	if (hdr_end > dreq->io_start)
 143		dreq_len = hdr_end - dreq->io_start;
 144
 145	nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 
 
 
 
 
 
 
 
 
 
 
 
 146
 147	if (dreq_len > dreq->max_count)
 148		dreq_len = dreq->max_count;
 149
 150	if (dreq->count < dreq_len)
 151		dreq->count = dreq_len;
 152}
 153
 154/*
 155 * nfs_direct_select_verf - select the right verifier
 156 * @dreq - direct request possibly spanning multiple servers
 157 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 158 * @commit_idx - commit bucket index for the DS
 159 *
 160 * returns the correct verifier to use given the role of the server
 161 */
 162static struct nfs_writeverf *
 163nfs_direct_select_verf(struct nfs_direct_req *dreq,
 164		       struct nfs_client *ds_clp,
 165		       int commit_idx)
 166{
 167	struct nfs_writeverf *verfp = &dreq->verf;
 168
 169#ifdef CONFIG_NFS_V4_1
 170	/*
 171	 * pNFS is in use, use the DS verf except commit_through_mds is set
 172	 * for layout segment where nbuckets is zero.
 173	 */
 174	if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 175		if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 176			verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 177		else
 178			WARN_ON_ONCE(1);
 179	}
 180#endif
 181	return verfp;
 182}
 183
 184
 185/*
 186 * nfs_direct_set_hdr_verf - set the write/commit verifier
 187 * @dreq - direct request possibly spanning multiple servers
 188 * @hdr - pageio header to validate against previously seen verfs
 189 *
 190 * Set the server's (MDS or DS) "seen" verifier
 191 */
 192static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 193				    struct nfs_pgio_header *hdr)
 194{
 195	struct nfs_writeverf *verfp;
 196
 197	verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 198	WARN_ON_ONCE(verfp->committed >= 0);
 199	memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 200	WARN_ON_ONCE(verfp->committed < 0);
 201}
 202
 203static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
 204		const struct nfs_writeverf *v2)
 205{
 206	return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
 207}
 208
 209/*
 210 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 211 * @dreq - direct request possibly spanning multiple servers
 212 * @hdr - pageio header to validate against previously seen verf
 213 *
 214 * set the server's "seen" verf if not initialized.
 215 * returns result of comparison between @hdr->verf and the "seen"
 216 * verf of the server used by @hdr (DS or MDS)
 217 */
 218static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 219					  struct nfs_pgio_header *hdr)
 220{
 221	struct nfs_writeverf *verfp;
 222
 223	verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 224	if (verfp->committed < 0) {
 225		nfs_direct_set_hdr_verf(dreq, hdr);
 226		return 0;
 227	}
 228	return nfs_direct_cmp_verf(verfp, &hdr->verf);
 229}
 230
 231/*
 232 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 233 * @dreq - direct request possibly spanning multiple servers
 234 * @data - commit data to validate against previously seen verf
 235 *
 236 * returns result of comparison between @data->verf and the verf of
 237 * the server used by @data (DS or MDS)
 238 */
 239static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 240					   struct nfs_commit_data *data)
 241{
 242	struct nfs_writeverf *verfp;
 243
 244	verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 245					 data->ds_commit_index);
 246
 247	/* verifier not set so always fail */
 248	if (verfp->committed < 0)
 249		return 1;
 250
 251	return nfs_direct_cmp_verf(verfp, &data->verf);
 252}
 253
 254/**
 255 * nfs_direct_IO - NFS address space operation for direct I/O
 256 * @iocb: target I/O control block
 257 * @iter: I/O buffer
 258 *
 259 * The presence of this routine in the address space ops vector means
 260 * the NFS client supports direct I/O. However, for most direct IO, we
 261 * shunt off direct read and write requests before the VFS gets them,
 262 * so this method is only ever called for swap.
 263 */
 264ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 265{
 266	struct inode *inode = iocb->ki_filp->f_mapping->host;
 267
 268	/* we only support swap file calling nfs_direct_IO */
 269	if (!IS_SWAPFILE(inode))
 270		return 0;
 271
 272	VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 273
 274	if (iov_iter_rw(iter) == READ)
 275		return nfs_file_direct_read(iocb, iter);
 276	return nfs_file_direct_write(iocb, iter);
 277}
 278
 279static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 280{
 281	unsigned int i;
 282	for (i = 0; i < npages; i++)
 283		put_page(pages[i]);
 284}
 285
 286void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 287			      struct nfs_direct_req *dreq)
 288{
 289	cinfo->inode = dreq->inode;
 290	cinfo->mds = &dreq->mds_cinfo;
 291	cinfo->ds = &dreq->ds_cinfo;
 292	cinfo->dreq = dreq;
 293	cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 294}
 295
 
 
 
 
 
 
 
 
 
 
 
 
 296static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 297{
 298	struct nfs_direct_req *dreq;
 299
 300	dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 301	if (!dreq)
 302		return NULL;
 303
 304	kref_init(&dreq->kref);
 305	kref_get(&dreq->kref);
 306	init_completion(&dreq->completion);
 307	INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 308	dreq->verf.committed = NFS_INVALID_STABLE_HOW;	/* not set yet */
 309	INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 
 310	spin_lock_init(&dreq->lock);
 311
 312	return dreq;
 313}
 314
 315static void nfs_direct_req_free(struct kref *kref)
 316{
 317	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 318
 319	nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 320	if (dreq->l_ctx != NULL)
 321		nfs_put_lock_context(dreq->l_ctx);
 322	if (dreq->ctx != NULL)
 323		put_nfs_open_context(dreq->ctx);
 324	kmem_cache_free(nfs_direct_cachep, dreq);
 325}
 326
 327static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 328{
 329	kref_put(&dreq->kref, nfs_direct_req_free);
 330}
 331
 332ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 333{
 334	return dreq->bytes_left;
 335}
 336EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 337
 338/*
 339 * Collects and returns the final error value/byte-count.
 340 */
 341static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 342{
 343	ssize_t result = -EIOCBQUEUED;
 344
 345	/* Async requests don't wait here */
 346	if (dreq->iocb)
 347		goto out;
 348
 349	result = wait_for_completion_killable(&dreq->completion);
 350
 351	if (!result) {
 352		result = dreq->count;
 353		WARN_ON_ONCE(dreq->count < 0);
 354	}
 355	if (!result)
 356		result = dreq->error;
 357
 358out:
 359	return (ssize_t) result;
 360}
 361
 362/*
 363 * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 364 * the iocb is still valid here if this is a synchronous request.
 365 */
 366static void nfs_direct_complete(struct nfs_direct_req *dreq)
 367{
 368	struct inode *inode = dreq->inode;
 369
 370	inode_dio_end(inode);
 371
 372	if (dreq->iocb) {
 373		long res = (long) dreq->error;
 374		if (dreq->count != 0) {
 375			res = (long) dreq->count;
 376			WARN_ON_ONCE(dreq->count < 0);
 377		}
 378		dreq->iocb->ki_complete(dreq->iocb, res, 0);
 379	}
 380
 381	complete(&dreq->completion);
 382
 383	nfs_direct_req_release(dreq);
 384}
 385
 386static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 387{
 388	unsigned long bytes = 0;
 389	struct nfs_direct_req *dreq = hdr->dreq;
 390
 391	spin_lock(&dreq->lock);
 392	if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 393		spin_unlock(&dreq->lock);
 394		goto out_put;
 395	}
 396
 397	nfs_direct_count_bytes(dreq, hdr);
 
 
 
 
 
 398	spin_unlock(&dreq->lock);
 399
 400	while (!list_empty(&hdr->pages)) {
 401		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 402		struct page *page = req->wb_page;
 403
 404		if (!PageCompound(page) && bytes < hdr->good_bytes &&
 405		    (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 406			set_page_dirty(page);
 407		bytes += req->wb_bytes;
 408		nfs_list_remove_request(req);
 409		nfs_release_request(req);
 410	}
 411out_put:
 412	if (put_dreq(dreq))
 413		nfs_direct_complete(dreq);
 414	hdr->release(hdr);
 415}
 416
 417static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 418{
 419	struct nfs_page *req;
 420
 421	while (!list_empty(head)) {
 422		req = nfs_list_entry(head->next);
 423		nfs_list_remove_request(req);
 424		nfs_release_request(req);
 425	}
 426}
 427
 428static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 429{
 430	get_dreq(hdr->dreq);
 431}
 432
 433static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 434	.error_cleanup = nfs_read_sync_pgio_error,
 435	.init_hdr = nfs_direct_pgio_init,
 436	.completion = nfs_direct_read_completion,
 437};
 438
 439/*
 440 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 441 * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 442 * bail and stop sending more reads.  Read length accounting is
 443 * handled automatically by nfs_direct_read_result().  Otherwise, if
 444 * no requests have been sent, just return an error.
 445 */
 446
 447static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 448					      struct iov_iter *iter,
 449					      loff_t pos)
 450{
 451	struct nfs_pageio_descriptor desc;
 452	struct inode *inode = dreq->inode;
 453	ssize_t result = -EINVAL;
 454	size_t requested_bytes = 0;
 455	size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 456
 457	nfs_pageio_init_read(&desc, dreq->inode, false,
 458			     &nfs_direct_read_completion_ops);
 459	get_dreq(dreq);
 460	desc.pg_dreq = dreq;
 461	inode_dio_begin(inode);
 462
 463	while (iov_iter_count(iter)) {
 464		struct page **pagevec;
 465		size_t bytes;
 466		size_t pgbase;
 467		unsigned npages, i;
 468
 469		result = iov_iter_get_pages_alloc(iter, &pagevec, 
 470						  rsize, &pgbase);
 471		if (result < 0)
 472			break;
 473	
 474		bytes = result;
 475		iov_iter_advance(iter, bytes);
 476		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 477		for (i = 0; i < npages; i++) {
 478			struct nfs_page *req;
 479			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 480			/* XXX do we need to do the eof zeroing found in async_filler? */
 481			req = nfs_create_request(dreq->ctx, pagevec[i],
 482						 pgbase, req_len);
 483			if (IS_ERR(req)) {
 484				result = PTR_ERR(req);
 485				break;
 486			}
 487			req->wb_index = pos >> PAGE_SHIFT;
 488			req->wb_offset = pos & ~PAGE_MASK;
 489			if (!nfs_pageio_add_request(&desc, req)) {
 490				result = desc.pg_error;
 491				nfs_release_request(req);
 492				break;
 493			}
 494			pgbase = 0;
 495			bytes -= req_len;
 496			requested_bytes += req_len;
 497			pos += req_len;
 498			dreq->bytes_left -= req_len;
 499		}
 500		nfs_direct_release_pages(pagevec, npages);
 501		kvfree(pagevec);
 502		if (result < 0)
 503			break;
 504	}
 505
 506	nfs_pageio_complete(&desc);
 507
 508	/*
 509	 * If no bytes were started, return the error, and let the
 510	 * generic layer handle the completion.
 511	 */
 512	if (requested_bytes == 0) {
 513		inode_dio_end(inode);
 514		nfs_direct_req_release(dreq);
 515		return result < 0 ? result : -EIO;
 516	}
 517
 518	if (put_dreq(dreq))
 519		nfs_direct_complete(dreq);
 520	return requested_bytes;
 521}
 522
 523/**
 524 * nfs_file_direct_read - file direct read operation for NFS files
 525 * @iocb: target I/O control block
 526 * @iter: vector of user buffers into which to read data
 527 *
 528 * We use this function for direct reads instead of calling
 529 * generic_file_aio_read() in order to avoid gfar's check to see if
 530 * the request starts before the end of the file.  For that check
 531 * to work, we must generate a GETATTR before each direct read, and
 532 * even then there is a window between the GETATTR and the subsequent
 533 * READ where the file size could change.  Our preference is simply
 534 * to do all reads the application wants, and the server will take
 535 * care of managing the end of file boundary.
 536 *
 537 * This function also eliminates unnecessarily updating the file's
 538 * atime locally, as the NFS server sets the file's atime, and this
 539 * client must read the updated atime from the server back into its
 540 * cache.
 541 */
 542ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
 543{
 544	struct file *file = iocb->ki_filp;
 545	struct address_space *mapping = file->f_mapping;
 546	struct inode *inode = mapping->host;
 547	struct nfs_direct_req *dreq;
 548	struct nfs_lock_context *l_ctx;
 549	ssize_t result = -EINVAL, requested;
 550	size_t count = iov_iter_count(iter);
 551	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 552
 553	dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 554		file, count, (long long) iocb->ki_pos);
 555
 556	result = 0;
 557	if (!count)
 558		goto out;
 559
 560	task_io_account_read(count);
 561
 562	result = -ENOMEM;
 563	dreq = nfs_direct_req_alloc();
 564	if (dreq == NULL)
 565		goto out;
 566
 567	dreq->inode = inode;
 568	dreq->bytes_left = dreq->max_count = count;
 569	dreq->io_start = iocb->ki_pos;
 570	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 571	l_ctx = nfs_get_lock_context(dreq->ctx);
 572	if (IS_ERR(l_ctx)) {
 573		result = PTR_ERR(l_ctx);
 574		goto out_release;
 575	}
 576	dreq->l_ctx = l_ctx;
 577	if (!is_sync_kiocb(iocb))
 578		dreq->iocb = iocb;
 579
 580	if (iter_is_iovec(iter))
 581		dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 582
 583	nfs_start_io_direct(inode);
 584
 585	NFS_I(inode)->read_io += count;
 586	requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 587
 588	nfs_end_io_direct(inode);
 589
 590	if (requested > 0) {
 591		result = nfs_direct_wait(dreq);
 592		if (result > 0) {
 593			requested -= result;
 594			iocb->ki_pos += result;
 595		}
 596		iov_iter_revert(iter, requested);
 597	} else {
 598		result = requested;
 599	}
 600
 601out_release:
 602	nfs_direct_req_release(dreq);
 603out:
 604	return result;
 605}
 606
 607static void
 608nfs_direct_write_scan_commit_list(struct inode *inode,
 609				  struct list_head *list,
 610				  struct nfs_commit_info *cinfo)
 611{
 612	mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 613#ifdef CONFIG_NFS_V4_1
 614	if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 615		NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 616#endif
 617	nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 618	mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 619}
 620
 621static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 622{
 623	struct nfs_pageio_descriptor desc;
 624	struct nfs_page *req, *tmp;
 625	LIST_HEAD(reqs);
 626	struct nfs_commit_info cinfo;
 627	LIST_HEAD(failed);
 
 628
 629	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 630	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 631
 632	dreq->count = 0;
 633	dreq->max_count = 0;
 634	list_for_each_entry(req, &reqs, wb_list)
 635		dreq->max_count += req->wb_bytes;
 636	dreq->verf.committed = NFS_INVALID_STABLE_HOW;
 637	nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 
 
 638	get_dreq(dreq);
 639
 640	nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 641			      &nfs_direct_write_completion_ops);
 642	desc.pg_dreq = dreq;
 643
 
 
 
 
 
 
 
 644	list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 645		/* Bump the transmission count */
 646		req->wb_nio++;
 647		if (!nfs_pageio_add_request(&desc, req)) {
 648			nfs_list_move_request(req, &failed);
 
 649			spin_lock(&cinfo.inode->i_lock);
 650			dreq->flags = 0;
 651			if (desc.pg_error < 0)
 652				dreq->error = desc.pg_error;
 653			else
 654				dreq->error = -EIO;
 655			spin_unlock(&cinfo.inode->i_lock);
 656		}
 657		nfs_release_request(req);
 658	}
 659	nfs_pageio_complete(&desc);
 660
 
 661	while (!list_empty(&failed)) {
 662		req = nfs_list_entry(failed.next);
 663		nfs_list_remove_request(req);
 664		nfs_unlock_and_release_request(req);
 665	}
 666
 667	if (put_dreq(dreq))
 668		nfs_direct_write_complete(dreq);
 669}
 670
 671static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 672{
 673	struct nfs_direct_req *dreq = data->dreq;
 674	struct nfs_commit_info cinfo;
 675	struct nfs_page *req;
 676	int status = data->task.tk_status;
 677
 678	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 679	if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
 680		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 681
 682	while (!list_empty(&data->pages)) {
 683		req = nfs_list_entry(data->pages.next);
 684		nfs_list_remove_request(req);
 685		if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 686			/*
 687			 * Despite the reboot, the write was successful,
 688			 * so reset wb_nio.
 689			 */
 690			req->wb_nio = 0;
 691			/* Note the rewrite will go through mds */
 692			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 693		} else
 694			nfs_release_request(req);
 695		nfs_unlock_and_release_request(req);
 696	}
 697
 698	if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 699		nfs_direct_write_complete(dreq);
 700}
 701
 702static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 703		struct nfs_page *req)
 704{
 705	struct nfs_direct_req *dreq = cinfo->dreq;
 706
 707	spin_lock(&dreq->lock);
 708	dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 709	spin_unlock(&dreq->lock);
 710	nfs_mark_request_commit(req, NULL, cinfo, 0);
 711}
 712
 713static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 714	.completion = nfs_direct_commit_complete,
 715	.resched_write = nfs_direct_resched_write,
 716};
 717
 718static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 719{
 720	int res;
 721	struct nfs_commit_info cinfo;
 722	LIST_HEAD(mds_list);
 723
 724	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 725	nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 726	res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 727	if (res < 0) /* res == -ENOMEM */
 728		nfs_direct_write_reschedule(dreq);
 729}
 730
 731static void nfs_direct_write_schedule_work(struct work_struct *work)
 732{
 733	struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 734	int flags = dreq->flags;
 735
 736	dreq->flags = 0;
 737	switch (flags) {
 738		case NFS_ODIRECT_DO_COMMIT:
 739			nfs_direct_commit_schedule(dreq);
 740			break;
 741		case NFS_ODIRECT_RESCHED_WRITES:
 742			nfs_direct_write_reschedule(dreq);
 743			break;
 744		default:
 745			nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 746			nfs_direct_complete(dreq);
 747	}
 748}
 749
 750static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 751{
 752	queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 753}
 754
 755static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 756{
 757	struct nfs_direct_req *dreq = hdr->dreq;
 758	struct nfs_commit_info cinfo;
 759	bool request_commit = false;
 760	struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 761
 
 
 
 762	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 763
 764	spin_lock(&dreq->lock);
 765	if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 766		spin_unlock(&dreq->lock);
 767		goto out_put;
 768	}
 769
 770	nfs_direct_count_bytes(dreq, hdr);
 771	if (hdr->good_bytes != 0) {
 
 
 772		if (nfs_write_need_commit(hdr)) {
 773			if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 774				request_commit = true;
 775			else if (dreq->flags == 0) {
 776				nfs_direct_set_hdr_verf(dreq, hdr);
 777				request_commit = true;
 778				dreq->flags = NFS_ODIRECT_DO_COMMIT;
 779			} else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 780				request_commit = true;
 781				if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 782					dreq->flags =
 783						NFS_ODIRECT_RESCHED_WRITES;
 784			}
 785		}
 786	}
 787	spin_unlock(&dreq->lock);
 788
 789	while (!list_empty(&hdr->pages)) {
 790
 791		req = nfs_list_entry(hdr->pages.next);
 792		nfs_list_remove_request(req);
 793		if (request_commit) {
 794			kref_get(&req->wb_kref);
 795			nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 796				hdr->ds_commit_idx);
 797		}
 798		nfs_unlock_and_release_request(req);
 799	}
 800
 801out_put:
 802	if (put_dreq(dreq))
 803		nfs_direct_write_complete(dreq);
 804	hdr->release(hdr);
 805}
 806
 807static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 808{
 809	struct nfs_page *req;
 810
 811	while (!list_empty(head)) {
 812		req = nfs_list_entry(head->next);
 813		nfs_list_remove_request(req);
 814		nfs_unlock_and_release_request(req);
 815	}
 816}
 817
 818static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 819{
 820	struct nfs_direct_req *dreq = hdr->dreq;
 821
 822	spin_lock(&dreq->lock);
 823	if (dreq->error == 0) {
 824		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 825		/* fake unstable write to let common nfs resend pages */
 826		hdr->verf.committed = NFS_UNSTABLE;
 827		hdr->good_bytes = hdr->args.count;
 828	}
 829	spin_unlock(&dreq->lock);
 830}
 831
 832static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 833	.error_cleanup = nfs_write_sync_pgio_error,
 834	.init_hdr = nfs_direct_pgio_init,
 835	.completion = nfs_direct_write_completion,
 836	.reschedule_io = nfs_direct_write_reschedule_io,
 837};
 838
 839
 840/*
 841 * NB: Return the value of the first error return code.  Subsequent
 842 *     errors after the first one are ignored.
 843 */
 844/*
 845 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 846 * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 847 * bail and stop sending more writes.  Write length accounting is
 848 * handled automatically by nfs_direct_write_result().  Otherwise, if
 849 * no requests have been sent, just return an error.
 850 */
 851static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 852					       struct iov_iter *iter,
 853					       loff_t pos)
 854{
 855	struct nfs_pageio_descriptor desc;
 856	struct inode *inode = dreq->inode;
 857	ssize_t result = 0;
 858	size_t requested_bytes = 0;
 859	size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 860
 861	nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 862			      &nfs_direct_write_completion_ops);
 863	desc.pg_dreq = dreq;
 864	get_dreq(dreq);
 865	inode_dio_begin(inode);
 866
 867	NFS_I(inode)->write_io += iov_iter_count(iter);
 868	while (iov_iter_count(iter)) {
 869		struct page **pagevec;
 870		size_t bytes;
 871		size_t pgbase;
 872		unsigned npages, i;
 873
 874		result = iov_iter_get_pages_alloc(iter, &pagevec, 
 875						  wsize, &pgbase);
 876		if (result < 0)
 877			break;
 878
 879		bytes = result;
 880		iov_iter_advance(iter, bytes);
 881		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 882		for (i = 0; i < npages; i++) {
 883			struct nfs_page *req;
 884			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 885
 886			req = nfs_create_request(dreq->ctx, pagevec[i],
 887						 pgbase, req_len);
 888			if (IS_ERR(req)) {
 889				result = PTR_ERR(req);
 890				break;
 891			}
 892
 
 893			if (desc.pg_error < 0) {
 894				nfs_free_request(req);
 895				result = desc.pg_error;
 896				break;
 897			}
 898
 899			nfs_lock_request(req);
 900			req->wb_index = pos >> PAGE_SHIFT;
 901			req->wb_offset = pos & ~PAGE_MASK;
 902			if (!nfs_pageio_add_request(&desc, req)) {
 903				result = desc.pg_error;
 904				nfs_unlock_and_release_request(req);
 905				break;
 906			}
 907			pgbase = 0;
 908			bytes -= req_len;
 909			requested_bytes += req_len;
 910			pos += req_len;
 911			dreq->bytes_left -= req_len;
 912		}
 913		nfs_direct_release_pages(pagevec, npages);
 914		kvfree(pagevec);
 915		if (result < 0)
 916			break;
 917	}
 918	nfs_pageio_complete(&desc);
 919
 920	/*
 921	 * If no bytes were started, return the error, and let the
 922	 * generic layer handle the completion.
 923	 */
 924	if (requested_bytes == 0) {
 925		inode_dio_end(inode);
 926		nfs_direct_req_release(dreq);
 927		return result < 0 ? result : -EIO;
 928	}
 929
 930	if (put_dreq(dreq))
 931		nfs_direct_write_complete(dreq);
 932	return requested_bytes;
 933}
 934
 935/**
 936 * nfs_file_direct_write - file direct write operation for NFS files
 937 * @iocb: target I/O control block
 938 * @iter: vector of user buffers from which to write data
 939 *
 940 * We use this function for direct writes instead of calling
 941 * generic_file_aio_write() in order to avoid taking the inode
 942 * semaphore and updating the i_size.  The NFS server will set
 943 * the new i_size and this client must read the updated size
 944 * back into its cache.  We let the server do generic write
 945 * parameter checking and report problems.
 946 *
 947 * We eliminate local atime updates, see direct read above.
 948 *
 949 * We avoid unnecessary page cache invalidations for normal cached
 950 * readers of this file.
 951 *
 952 * Note that O_APPEND is not supported for NFS direct writes, as there
 953 * is no atomic O_APPEND write facility in the NFS protocol.
 954 */
 955ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 956{
 957	ssize_t result = -EINVAL, requested;
 958	size_t count;
 959	struct file *file = iocb->ki_filp;
 960	struct address_space *mapping = file->f_mapping;
 961	struct inode *inode = mapping->host;
 962	struct nfs_direct_req *dreq;
 963	struct nfs_lock_context *l_ctx;
 964	loff_t pos, end;
 965
 966	dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 967		file, iov_iter_count(iter), (long long) iocb->ki_pos);
 968
 969	result = generic_write_checks(iocb, iter);
 970	if (result <= 0)
 971		return result;
 972	count = result;
 973	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 974
 975	pos = iocb->ki_pos;
 976	end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 977
 978	task_io_account_write(count);
 979
 980	result = -ENOMEM;
 981	dreq = nfs_direct_req_alloc();
 982	if (!dreq)
 983		goto out;
 984
 985	dreq->inode = inode;
 986	dreq->bytes_left = dreq->max_count = count;
 987	dreq->io_start = pos;
 988	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 989	l_ctx = nfs_get_lock_context(dreq->ctx);
 990	if (IS_ERR(l_ctx)) {
 991		result = PTR_ERR(l_ctx);
 992		goto out_release;
 993	}
 994	dreq->l_ctx = l_ctx;
 995	if (!is_sync_kiocb(iocb))
 996		dreq->iocb = iocb;
 997
 998	nfs_start_io_direct(inode);
 999
1000	requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1001
1002	if (mapping->nrpages) {
1003		invalidate_inode_pages2_range(mapping,
1004					      pos >> PAGE_SHIFT, end);
1005	}
1006
1007	nfs_end_io_direct(inode);
1008
1009	if (requested > 0) {
1010		result = nfs_direct_wait(dreq);
1011		if (result > 0) {
1012			requested -= result;
1013			iocb->ki_pos = pos + result;
1014			/* XXX: should check the generic_write_sync retval */
1015			generic_write_sync(iocb, result);
1016		}
1017		iov_iter_revert(iter, requested);
1018	} else {
1019		result = requested;
1020	}
1021out_release:
1022	nfs_direct_req_release(dreq);
1023out:
1024	return result;
1025}
1026
1027/**
1028 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1029 *
1030 */
1031int __init nfs_init_directcache(void)
1032{
1033	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1034						sizeof(struct nfs_direct_req),
1035						0, (SLAB_RECLAIM_ACCOUNT|
1036							SLAB_MEM_SPREAD),
1037						NULL);
1038	if (nfs_direct_cachep == NULL)
1039		return -ENOMEM;
1040
1041	return 0;
1042}
1043
1044/**
1045 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1046 *
1047 */
1048void nfs_destroy_directcache(void)
1049{
1050	kmem_cache_destroy(nfs_direct_cachep);
1051}
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}