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}
v6.13.7
   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 "delegation.h"
  60#include "internal.h"
  61#include "iostat.h"
  62#include "pnfs.h"
  63#include "fscache.h"
  64#include "nfstrace.h"
  65
  66#define NFSDBG_FACILITY		NFSDBG_VFS
  67
  68static struct kmem_cache *nfs_direct_cachep;
  69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  70static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  71static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  72static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
  73static void nfs_direct_write_schedule_work(struct work_struct *work);
  74
  75static inline void get_dreq(struct nfs_direct_req *dreq)
  76{
  77	atomic_inc(&dreq->io_count);
  78}
  79
  80static inline int put_dreq(struct nfs_direct_req *dreq)
  81{
  82	return atomic_dec_and_test(&dreq->io_count);
  83}
  84
  85static void
  86nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
  87			    const struct nfs_pgio_header *hdr,
  88			    ssize_t dreq_len)
  89{
  90	if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
  91	      test_bit(NFS_IOHDR_EOF, &hdr->flags)))
  92		return;
  93	if (dreq->max_count >= dreq_len) {
  94		dreq->max_count = dreq_len;
  95		if (dreq->count > dreq_len)
  96			dreq->count = dreq_len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  97	}
  98
  99	if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && !dreq->error)
 100		dreq->error = hdr->error;
 101}
 102
 103static void
 104nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 105		       const struct nfs_pgio_header *hdr)
 
 
 
 
 
 
 
 
 
 106{
 107	loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 108	ssize_t dreq_len = 0;
 109
 110	if (hdr_end > dreq->io_start)
 111		dreq_len = hdr_end - dreq->io_start;
 
 
 
 
 
 
 
 
 
 
 
 
 112
 113	nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 114
 115	if (dreq_len > dreq->max_count)
 116		dreq_len = dreq->max_count;
 
 
 
 
 
 
 
 
 
 117
 118	if (dreq->count < dreq_len)
 119		dreq->count = dreq_len;
 
 
 120}
 121
 122static void nfs_direct_truncate_request(struct nfs_direct_req *dreq,
 123					struct nfs_page *req)
 124{
 125	loff_t offs = req_offset(req);
 126	size_t req_start = (size_t)(offs - dreq->io_start);
 127
 128	if (req_start < dreq->max_count)
 129		dreq->max_count = req_start;
 130	if (req_start < dreq->count)
 131		dreq->count = req_start;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 132}
 133
 134static void nfs_direct_file_adjust_size_locked(struct inode *inode,
 135					       loff_t offset, size_t count)
 
 
 
 
 
 
 
 
 136{
 137	loff_t newsize = offset + (loff_t)count;
 138	loff_t oldsize = i_size_read(inode);
 139
 140	if (newsize > oldsize) {
 141		i_size_write(inode, newsize);
 142		NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 143		trace_nfs_size_grow(inode, newsize);
 144		nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 145	}
 
 
 146}
 147
 148/**
 149 * nfs_swap_rw - NFS address space operation for swap I/O
 150 * @iocb: target I/O control block
 151 * @iter: I/O buffer
 152 *
 153 * Perform IO to the swap-file.  This is much like direct IO.
 
 
 
 154 */
 155int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
 156{
 157	ssize_t ret;
 
 
 
 
 
 
 158
 159	if (iov_iter_rw(iter) == READ)
 160		ret = nfs_file_direct_read(iocb, iter, true);
 161	else
 162		ret = nfs_file_direct_write(iocb, iter, true);
 163	if (ret < 0)
 164		return ret;
 165	return 0;
 166}
 167
 168static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 169{
 170	unsigned int i;
 171	for (i = 0; i < npages; i++)
 172		put_page(pages[i]);
 173}
 174
 175void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 176			      struct nfs_direct_req *dreq)
 177{
 178	cinfo->inode = dreq->inode;
 179	cinfo->mds = &dreq->mds_cinfo;
 180	cinfo->ds = &dreq->ds_cinfo;
 181	cinfo->dreq = dreq;
 182	cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 183}
 184
 
 
 
 
 
 
 
 
 
 
 
 
 185static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 186{
 187	struct nfs_direct_req *dreq;
 188
 189	dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 190	if (!dreq)
 191		return NULL;
 192
 193	kref_init(&dreq->kref);
 194	kref_get(&dreq->kref);
 195	init_completion(&dreq->completion);
 196	INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 197	pnfs_init_ds_commit_info(&dreq->ds_cinfo);
 198	INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 
 199	spin_lock_init(&dreq->lock);
 200
 201	return dreq;
 202}
 203
 204static void nfs_direct_req_free(struct kref *kref)
 205{
 206	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 207
 208	pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
 209	if (dreq->l_ctx != NULL)
 210		nfs_put_lock_context(dreq->l_ctx);
 211	if (dreq->ctx != NULL)
 212		put_nfs_open_context(dreq->ctx);
 213	kmem_cache_free(nfs_direct_cachep, dreq);
 214}
 215
 216static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 217{
 218	kref_put(&dreq->kref, nfs_direct_req_free);
 219}
 220
 221ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq, loff_t offset)
 222{
 223	loff_t start = offset - dreq->io_start;
 224	return dreq->max_count - start;
 225}
 226EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 227
 228/*
 229 * Collects and returns the final error value/byte-count.
 230 */
 231static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 232{
 233	ssize_t result = -EIOCBQUEUED;
 234
 235	/* Async requests don't wait here */
 236	if (dreq->iocb)
 237		goto out;
 238
 239	result = wait_for_completion_killable(&dreq->completion);
 240
 241	if (!result) {
 242		result = dreq->count;
 243		WARN_ON_ONCE(dreq->count < 0);
 244	}
 245	if (!result)
 246		result = dreq->error;
 247
 248out:
 249	return (ssize_t) result;
 250}
 251
 252/*
 253 * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 254 * the iocb is still valid here if this is a synchronous request.
 255 */
 256static void nfs_direct_complete(struct nfs_direct_req *dreq)
 257{
 258	struct inode *inode = dreq->inode;
 259
 260	inode_dio_end(inode);
 261
 262	if (dreq->iocb) {
 263		long res = (long) dreq->error;
 264		if (dreq->count != 0) {
 265			res = (long) dreq->count;
 266			WARN_ON_ONCE(dreq->count < 0);
 267		}
 268		dreq->iocb->ki_complete(dreq->iocb, res);
 269	}
 270
 271	complete(&dreq->completion);
 272
 273	nfs_direct_req_release(dreq);
 274}
 275
 276static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 277{
 278	unsigned long bytes = 0;
 279	struct nfs_direct_req *dreq = hdr->dreq;
 280
 
 
 
 281	spin_lock(&dreq->lock);
 282	if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 283		spin_unlock(&dreq->lock);
 284		goto out_put;
 285	}
 286
 287	nfs_direct_count_bytes(dreq, hdr);
 288	spin_unlock(&dreq->lock);
 289
 290	nfs_update_delegated_atime(dreq->inode);
 291
 292	while (!list_empty(&hdr->pages)) {
 293		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 294		struct page *page = req->wb_page;
 295
 296		if (!PageCompound(page) && bytes < hdr->good_bytes &&
 297		    (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 298			set_page_dirty(page);
 299		bytes += req->wb_bytes;
 300		nfs_list_remove_request(req);
 301		nfs_release_request(req);
 302	}
 303out_put:
 304	if (put_dreq(dreq))
 305		nfs_direct_complete(dreq);
 306	hdr->release(hdr);
 307}
 308
 309static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 310{
 311	struct nfs_page *req;
 312
 313	while (!list_empty(head)) {
 314		req = nfs_list_entry(head->next);
 315		nfs_list_remove_request(req);
 316		nfs_release_request(req);
 317	}
 318}
 319
 320static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 321{
 322	get_dreq(hdr->dreq);
 323}
 324
 325static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 326	.error_cleanup = nfs_read_sync_pgio_error,
 327	.init_hdr = nfs_direct_pgio_init,
 328	.completion = nfs_direct_read_completion,
 329};
 330
 331/*
 332 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 333 * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 334 * bail and stop sending more reads.  Read length accounting is
 335 * handled automatically by nfs_direct_read_result().  Otherwise, if
 336 * no requests have been sent, just return an error.
 337 */
 338
 339static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 340					      struct iov_iter *iter,
 341					      loff_t pos)
 342{
 343	struct nfs_pageio_descriptor desc;
 344	struct inode *inode = dreq->inode;
 345	ssize_t result = -EINVAL;
 346	size_t requested_bytes = 0;
 347	size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 348
 349	nfs_pageio_init_read(&desc, dreq->inode, false,
 350			     &nfs_direct_read_completion_ops);
 351	get_dreq(dreq);
 352	desc.pg_dreq = dreq;
 353	inode_dio_begin(inode);
 354
 355	while (iov_iter_count(iter)) {
 356		struct page **pagevec;
 357		size_t bytes;
 358		size_t pgbase;
 359		unsigned npages, i;
 360
 361		result = iov_iter_get_pages_alloc2(iter, &pagevec,
 362						  rsize, &pgbase);
 363		if (result < 0)
 364			break;
 365	
 366		bytes = result;
 
 367		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 368		for (i = 0; i < npages; i++) {
 369			struct nfs_page *req;
 370			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 371			/* XXX do we need to do the eof zeroing found in async_filler? */
 372			req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 373							pgbase, pos, req_len);
 374			if (IS_ERR(req)) {
 375				result = PTR_ERR(req);
 376				break;
 377			}
 
 
 378			if (!nfs_pageio_add_request(&desc, req)) {
 379				result = desc.pg_error;
 380				nfs_release_request(req);
 381				break;
 382			}
 383			pgbase = 0;
 384			bytes -= req_len;
 385			requested_bytes += req_len;
 386			pos += req_len;
 
 387		}
 388		nfs_direct_release_pages(pagevec, npages);
 389		kvfree(pagevec);
 390		if (result < 0)
 391			break;
 392	}
 393
 394	nfs_pageio_complete(&desc);
 395
 396	/*
 397	 * If no bytes were started, return the error, and let the
 398	 * generic layer handle the completion.
 399	 */
 400	if (requested_bytes == 0) {
 401		inode_dio_end(inode);
 402		nfs_direct_req_release(dreq);
 403		return result < 0 ? result : -EIO;
 404	}
 405
 406	if (put_dreq(dreq))
 407		nfs_direct_complete(dreq);
 408	return requested_bytes;
 409}
 410
 411/**
 412 * nfs_file_direct_read - file direct read operation for NFS files
 413 * @iocb: target I/O control block
 414 * @iter: vector of user buffers into which to read data
 415 * @swap: flag indicating this is swap IO, not O_DIRECT IO
 416 *
 417 * We use this function for direct reads instead of calling
 418 * generic_file_aio_read() in order to avoid gfar's check to see if
 419 * the request starts before the end of the file.  For that check
 420 * to work, we must generate a GETATTR before each direct read, and
 421 * even then there is a window between the GETATTR and the subsequent
 422 * READ where the file size could change.  Our preference is simply
 423 * to do all reads the application wants, and the server will take
 424 * care of managing the end of file boundary.
 425 *
 426 * This function also eliminates unnecessarily updating the file's
 427 * atime locally, as the NFS server sets the file's atime, and this
 428 * client must read the updated atime from the server back into its
 429 * cache.
 430 */
 431ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 432			     bool swap)
 433{
 434	struct file *file = iocb->ki_filp;
 435	struct address_space *mapping = file->f_mapping;
 436	struct inode *inode = mapping->host;
 437	struct nfs_direct_req *dreq;
 438	struct nfs_lock_context *l_ctx;
 439	ssize_t result, requested;
 440	size_t count = iov_iter_count(iter);
 441	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 442
 443	dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 444		file, count, (long long) iocb->ki_pos);
 445
 446	result = 0;
 447	if (!count)
 448		goto out;
 449
 450	task_io_account_read(count);
 451
 452	result = -ENOMEM;
 453	dreq = nfs_direct_req_alloc();
 454	if (dreq == NULL)
 455		goto out;
 456
 457	dreq->inode = inode;
 458	dreq->max_count = count;
 459	dreq->io_start = iocb->ki_pos;
 460	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 461	l_ctx = nfs_get_lock_context(dreq->ctx);
 462	if (IS_ERR(l_ctx)) {
 463		result = PTR_ERR(l_ctx);
 464		nfs_direct_req_release(dreq);
 465		goto out_release;
 466	}
 467	dreq->l_ctx = l_ctx;
 468	if (!is_sync_kiocb(iocb))
 469		dreq->iocb = iocb;
 470
 471	if (user_backed_iter(iter))
 472		dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 473
 474	if (!swap) {
 475		result = nfs_start_io_direct(inode);
 476		if (result) {
 477			/* release the reference that would usually be
 478			 * consumed by nfs_direct_read_schedule_iovec()
 479			 */
 480			nfs_direct_req_release(dreq);
 481			goto out_release;
 482		}
 483	}
 484
 485	NFS_I(inode)->read_io += count;
 486	requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 487
 488	if (!swap)
 489		nfs_end_io_direct(inode);
 490
 491	if (requested > 0) {
 492		result = nfs_direct_wait(dreq);
 493		if (result > 0) {
 494			requested -= result;
 495			iocb->ki_pos += result;
 496		}
 497		iov_iter_revert(iter, requested);
 498	} else {
 499		result = requested;
 500	}
 501
 502out_release:
 503	nfs_direct_req_release(dreq);
 504out:
 505	return result;
 506}
 507
 508static void nfs_direct_add_page_head(struct list_head *list,
 509				     struct nfs_page *req)
 510{
 511	struct nfs_page *head = req->wb_head;
 512
 513	if (!list_empty(&head->wb_list) || !nfs_lock_request(head))
 514		return;
 515	if (!list_empty(&head->wb_list)) {
 516		nfs_unlock_request(head);
 517		return;
 518	}
 519	list_add(&head->wb_list, list);
 520	kref_get(&head->wb_kref);
 521	kref_get(&head->wb_kref);
 522}
 523
 524static void nfs_direct_join_group(struct list_head *list,
 525				  struct nfs_commit_info *cinfo,
 526				  struct inode *inode)
 527{
 528	struct nfs_page *req, *subreq;
 529
 530	list_for_each_entry(req, list, wb_list) {
 531		if (req->wb_head != req) {
 532			nfs_direct_add_page_head(&req->wb_list, req);
 533			continue;
 534		}
 535		subreq = req->wb_this_page;
 536		if (subreq == req)
 537			continue;
 538		do {
 539			/*
 540			 * Remove subrequests from this list before freeing
 541			 * them in the call to nfs_join_page_group().
 542			 */
 543			if (!list_empty(&subreq->wb_list)) {
 544				nfs_list_remove_request(subreq);
 545				nfs_release_request(subreq);
 546			}
 547		} while ((subreq = subreq->wb_this_page) != req);
 548		nfs_join_page_group(req, cinfo, inode);
 549	}
 550}
 551
 552static void
 553nfs_direct_write_scan_commit_list(struct inode *inode,
 554				  struct list_head *list,
 555				  struct nfs_commit_info *cinfo)
 556{
 557	mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 558	pnfs_recover_commit_reqs(list, cinfo);
 
 
 
 559	nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 560	mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 561}
 562
 563static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 564{
 565	struct nfs_pageio_descriptor desc;
 566	struct nfs_page *req;
 567	LIST_HEAD(reqs);
 568	struct nfs_commit_info cinfo;
 
 
 569
 570	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 571	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 572
 573	nfs_direct_join_group(&reqs, &cinfo, dreq->inode);
 574
 575	nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 
 
 576	get_dreq(dreq);
 577
 578	nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 579			      &nfs_direct_write_completion_ops);
 580	desc.pg_dreq = dreq;
 581
 582	while (!list_empty(&reqs)) {
 583		req = nfs_list_entry(reqs.next);
 584		/* Bump the transmission count */
 585		req->wb_nio++;
 
 
 
 
 586		if (!nfs_pageio_add_request(&desc, req)) {
 587			spin_lock(&dreq->lock);
 588			if (dreq->error < 0) {
 589				desc.pg_error = dreq->error;
 590			} else if (desc.pg_error != -EAGAIN) {
 591				dreq->flags = 0;
 592				if (!desc.pg_error)
 593					desc.pg_error = -EIO;
 594				dreq->error = desc.pg_error;
 595			} else
 596				dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 597			spin_unlock(&dreq->lock);
 598			break;
 599		}
 600		nfs_release_request(req);
 601	}
 602	nfs_pageio_complete(&desc);
 603
 604	while (!list_empty(&reqs)) {
 605		req = nfs_list_entry(reqs.next);
 
 606		nfs_list_remove_request(req);
 607		nfs_unlock_and_release_request(req);
 608		if (desc.pg_error == -EAGAIN) {
 609			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 610		} else {
 611			spin_lock(&dreq->lock);
 612			nfs_direct_truncate_request(dreq, req);
 613			spin_unlock(&dreq->lock);
 614			nfs_release_request(req);
 615		}
 616	}
 617
 618	if (put_dreq(dreq))
 619		nfs_direct_write_complete(dreq);
 620}
 621
 622static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 623{
 624	const struct nfs_writeverf *verf = data->res.verf;
 625	struct nfs_direct_req *dreq = data->dreq;
 626	struct nfs_commit_info cinfo;
 627	struct nfs_page *req;
 628	int status = data->task.tk_status;
 629
 630	trace_nfs_direct_commit_complete(dreq);
 631
 632	spin_lock(&dreq->lock);
 633	if (status < 0) {
 634		/* Errors in commit are fatal */
 635		dreq->error = status;
 636		dreq->flags = NFS_ODIRECT_DONE;
 637	} else {
 638		status = dreq->error;
 639	}
 640	spin_unlock(&dreq->lock);
 641
 642	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 
 
 643
 644	while (!list_empty(&data->pages)) {
 645		req = nfs_list_entry(data->pages.next);
 646		nfs_list_remove_request(req);
 647		if (status < 0) {
 648			spin_lock(&dreq->lock);
 649			nfs_direct_truncate_request(dreq, req);
 650			spin_unlock(&dreq->lock);
 651			nfs_release_request(req);
 652		} else if (!nfs_write_match_verf(verf, req)) {
 653			spin_lock(&dreq->lock);
 654			if (dreq->flags == 0)
 655				dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 656			spin_unlock(&dreq->lock);
 657			/*
 658			 * Despite the reboot, the write was successful,
 659			 * so reset wb_nio.
 660			 */
 661			req->wb_nio = 0;
 662			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 663		} else
 664			nfs_release_request(req);
 665		nfs_unlock_and_release_request(req);
 666	}
 667
 668	if (nfs_commit_end(cinfo.mds))
 669		nfs_direct_write_complete(dreq);
 670}
 671
 672static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 673		struct nfs_page *req)
 674{
 675	struct nfs_direct_req *dreq = cinfo->dreq;
 676
 677	trace_nfs_direct_resched_write(dreq);
 678
 679	spin_lock(&dreq->lock);
 680	if (dreq->flags != NFS_ODIRECT_DONE)
 681		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 682	spin_unlock(&dreq->lock);
 683	nfs_mark_request_commit(req, NULL, cinfo, 0);
 684}
 685
 686static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 687	.completion = nfs_direct_commit_complete,
 688	.resched_write = nfs_direct_resched_write,
 689};
 690
 691static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 692{
 693	int res;
 694	struct nfs_commit_info cinfo;
 695	LIST_HEAD(mds_list);
 696
 697	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 698	nfs_commit_begin(cinfo.mds);
 699	nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 700	res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 701	if (res < 0) { /* res == -ENOMEM */
 702		spin_lock(&dreq->lock);
 703		if (dreq->flags == 0)
 704			dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 705		spin_unlock(&dreq->lock);
 706	}
 707	if (nfs_commit_end(cinfo.mds))
 708		nfs_direct_write_complete(dreq);
 709}
 710
 711static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
 712{
 713	struct nfs_commit_info cinfo;
 714	struct nfs_page *req;
 715	LIST_HEAD(reqs);
 716
 717	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 718	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 719
 720	while (!list_empty(&reqs)) {
 721		req = nfs_list_entry(reqs.next);
 722		nfs_list_remove_request(req);
 723		nfs_direct_truncate_request(dreq, req);
 724		nfs_release_request(req);
 725		nfs_unlock_and_release_request(req);
 726	}
 727}
 728
 729static void nfs_direct_write_schedule_work(struct work_struct *work)
 730{
 731	struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 732	int flags = dreq->flags;
 733
 734	dreq->flags = 0;
 735	switch (flags) {
 736		case NFS_ODIRECT_DO_COMMIT:
 737			nfs_direct_commit_schedule(dreq);
 738			break;
 739		case NFS_ODIRECT_RESCHED_WRITES:
 740			nfs_direct_write_reschedule(dreq);
 741			break;
 742		default:
 743			nfs_direct_write_clear_reqs(dreq);
 744			nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 745			nfs_direct_complete(dreq);
 746	}
 747}
 748
 749static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 750{
 751	trace_nfs_direct_write_complete(dreq);
 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	struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 760	struct inode *inode = dreq->inode;
 761	int flags = NFS_ODIRECT_DONE;
 762
 763	trace_nfs_direct_write_completion(dreq);
 
 764
 765	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 766
 767	spin_lock(&dreq->lock);
 768	if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 769		spin_unlock(&dreq->lock);
 770		goto out_put;
 771	}
 772
 773	nfs_direct_count_bytes(dreq, hdr);
 774	if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags) &&
 775	    !test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 776		if (!dreq->flags)
 777			dreq->flags = NFS_ODIRECT_DO_COMMIT;
 778		flags = dreq->flags;
 
 
 
 
 
 
 
 
 
 
 
 
 779	}
 780	spin_unlock(&dreq->lock);
 781
 782	spin_lock(&inode->i_lock);
 783	nfs_direct_file_adjust_size_locked(inode, dreq->io_start, dreq->count);
 784	nfs_update_delegated_mtime_locked(dreq->inode);
 785	spin_unlock(&inode->i_lock);
 786
 787	while (!list_empty(&hdr->pages)) {
 788
 789		req = nfs_list_entry(hdr->pages.next);
 790		nfs_list_remove_request(req);
 791		if (flags == NFS_ODIRECT_DO_COMMIT) {
 792			kref_get(&req->wb_kref);
 793			memcpy(&req->wb_verf, &hdr->verf.verifier,
 794			       sizeof(req->wb_verf));
 795			nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 796				hdr->ds_commit_idx);
 797		} else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
 798			kref_get(&req->wb_kref);
 799			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 800		}
 801		nfs_unlock_and_release_request(req);
 802	}
 803
 804out_put:
 805	if (put_dreq(dreq))
 806		nfs_direct_write_complete(dreq);
 807	hdr->release(hdr);
 808}
 809
 810static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 811{
 812	struct nfs_page *req;
 813
 814	while (!list_empty(head)) {
 815		req = nfs_list_entry(head->next);
 816		nfs_list_remove_request(req);
 817		nfs_unlock_and_release_request(req);
 818	}
 819}
 820
 821static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 822{
 823	struct nfs_direct_req *dreq = hdr->dreq;
 824	struct nfs_page *req;
 825	struct nfs_commit_info cinfo;
 826
 827	trace_nfs_direct_write_reschedule_io(dreq);
 828
 829	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 830	spin_lock(&dreq->lock);
 831	if (dreq->error == 0)
 832		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 833	set_bit(NFS_IOHDR_REDO, &hdr->flags);
 
 
 
 834	spin_unlock(&dreq->lock);
 835	while (!list_empty(&hdr->pages)) {
 836		req = nfs_list_entry(hdr->pages.next);
 837		nfs_list_remove_request(req);
 838		nfs_unlock_request(req);
 839		nfs_mark_request_commit(req, NULL, &cinfo, 0);
 840	}
 841}
 842
 843static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 844	.error_cleanup = nfs_write_sync_pgio_error,
 845	.init_hdr = nfs_direct_pgio_init,
 846	.completion = nfs_direct_write_completion,
 847	.reschedule_io = nfs_direct_write_reschedule_io,
 848};
 849
 850
 851/*
 852 * NB: Return the value of the first error return code.  Subsequent
 853 *     errors after the first one are ignored.
 854 */
 855/*
 856 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 857 * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 858 * bail and stop sending more writes.  Write length accounting is
 859 * handled automatically by nfs_direct_write_result().  Otherwise, if
 860 * no requests have been sent, just return an error.
 861 */
 862static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 863					       struct iov_iter *iter,
 864					       loff_t pos, int ioflags)
 865{
 866	struct nfs_pageio_descriptor desc;
 867	struct inode *inode = dreq->inode;
 868	struct nfs_commit_info cinfo;
 869	ssize_t result = 0;
 870	size_t requested_bytes = 0;
 871	size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 872	bool defer = false;
 873
 874	trace_nfs_direct_write_schedule_iovec(dreq);
 875
 876	nfs_pageio_init_write(&desc, inode, ioflags, false,
 877			      &nfs_direct_write_completion_ops);
 878	desc.pg_dreq = dreq;
 879	get_dreq(dreq);
 880	inode_dio_begin(inode);
 881
 882	NFS_I(inode)->write_io += iov_iter_count(iter);
 883	while (iov_iter_count(iter)) {
 884		struct page **pagevec;
 885		size_t bytes;
 886		size_t pgbase;
 887		unsigned npages, i;
 888
 889		result = iov_iter_get_pages_alloc2(iter, &pagevec,
 890						  wsize, &pgbase);
 891		if (result < 0)
 892			break;
 893
 894		bytes = result;
 
 895		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 896		for (i = 0; i < npages; i++) {
 897			struct nfs_page *req;
 898			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 899
 900			req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 901							pgbase, pos, req_len);
 902			if (IS_ERR(req)) {
 903				result = PTR_ERR(req);
 904				break;
 905			}
 906
 
 907			if (desc.pg_error < 0) {
 908				nfs_free_request(req);
 909				result = desc.pg_error;
 910				break;
 911			}
 912
 913			pgbase = 0;
 914			bytes -= req_len;
 915			requested_bytes += req_len;
 916			pos += req_len;
 917
 918			if (defer) {
 919				nfs_mark_request_commit(req, NULL, &cinfo, 0);
 920				continue;
 921			}
 922
 923			nfs_lock_request(req);
 924			if (nfs_pageio_add_request(&desc, req))
 925				continue;
 926
 927			/* Exit on hard errors */
 928			if (desc.pg_error < 0 && desc.pg_error != -EAGAIN) {
 929				result = desc.pg_error;
 930				nfs_unlock_and_release_request(req);
 931				break;
 932			}
 933
 934			/* If the error is soft, defer remaining requests */
 935			nfs_init_cinfo_from_dreq(&cinfo, dreq);
 936			spin_lock(&dreq->lock);
 937			dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 938			spin_unlock(&dreq->lock);
 939			nfs_unlock_request(req);
 940			nfs_mark_request_commit(req, NULL, &cinfo, 0);
 941			desc.pg_error = 0;
 942			defer = true;
 943		}
 944		nfs_direct_release_pages(pagevec, npages);
 945		kvfree(pagevec);
 946		if (result < 0)
 947			break;
 948	}
 949	nfs_pageio_complete(&desc);
 950
 951	/*
 952	 * If no bytes were started, return the error, and let the
 953	 * generic layer handle the completion.
 954	 */
 955	if (requested_bytes == 0) {
 956		inode_dio_end(inode);
 957		nfs_direct_req_release(dreq);
 958		return result < 0 ? result : -EIO;
 959	}
 960
 961	if (put_dreq(dreq))
 962		nfs_direct_write_complete(dreq);
 963	return requested_bytes;
 964}
 965
 966/**
 967 * nfs_file_direct_write - file direct write operation for NFS files
 968 * @iocb: target I/O control block
 969 * @iter: vector of user buffers from which to write data
 970 * @swap: flag indicating this is swap IO, not O_DIRECT IO
 971 *
 972 * We use this function for direct writes instead of calling
 973 * generic_file_aio_write() in order to avoid taking the inode
 974 * semaphore and updating the i_size.  The NFS server will set
 975 * the new i_size and this client must read the updated size
 976 * back into its cache.  We let the server do generic write
 977 * parameter checking and report problems.
 978 *
 979 * We eliminate local atime updates, see direct read above.
 980 *
 981 * We avoid unnecessary page cache invalidations for normal cached
 982 * readers of this file.
 983 *
 984 * Note that O_APPEND is not supported for NFS direct writes, as there
 985 * is no atomic O_APPEND write facility in the NFS protocol.
 986 */
 987ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 988			      bool swap)
 989{
 990	ssize_t result, requested;
 991	size_t count;
 992	struct file *file = iocb->ki_filp;
 993	struct address_space *mapping = file->f_mapping;
 994	struct inode *inode = mapping->host;
 995	struct nfs_direct_req *dreq;
 996	struct nfs_lock_context *l_ctx;
 997	loff_t pos, end;
 998
 999	dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1000		file, iov_iter_count(iter), (long long) iocb->ki_pos);
1001
1002	if (swap)
1003		/* bypass generic checks */
1004		result =  iov_iter_count(iter);
1005	else
1006		result = generic_write_checks(iocb, iter);
1007	if (result <= 0)
1008		return result;
1009	count = result;
1010	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1011
1012	pos = iocb->ki_pos;
1013	end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1014
1015	task_io_account_write(count);
1016
1017	result = -ENOMEM;
1018	dreq = nfs_direct_req_alloc();
1019	if (!dreq)
1020		goto out;
1021
1022	dreq->inode = inode;
1023	dreq->max_count = count;
1024	dreq->io_start = pos;
1025	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1026	l_ctx = nfs_get_lock_context(dreq->ctx);
1027	if (IS_ERR(l_ctx)) {
1028		result = PTR_ERR(l_ctx);
1029		nfs_direct_req_release(dreq);
1030		goto out_release;
1031	}
1032	dreq->l_ctx = l_ctx;
1033	if (!is_sync_kiocb(iocb))
1034		dreq->iocb = iocb;
1035	pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
1036
1037	if (swap) {
1038		requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1039							    FLUSH_STABLE);
1040	} else {
1041		result = nfs_start_io_direct(inode);
1042		if (result) {
1043			/* release the reference that would usually be
1044			 * consumed by nfs_direct_write_schedule_iovec()
1045			 */
1046			nfs_direct_req_release(dreq);
1047			goto out_release;
1048		}
1049
1050		requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1051							    FLUSH_COND_STABLE);
1052
1053		if (mapping->nrpages) {
1054			invalidate_inode_pages2_range(mapping,
1055						      pos >> PAGE_SHIFT, end);
1056		}
1057
1058		nfs_end_io_direct(inode);
1059	}
1060
1061	if (requested > 0) {
1062		result = nfs_direct_wait(dreq);
1063		if (result > 0) {
1064			requested -= result;
1065			iocb->ki_pos = pos + result;
1066			/* XXX: should check the generic_write_sync retval */
1067			generic_write_sync(iocb, result);
1068		}
1069		iov_iter_revert(iter, requested);
1070	} else {
1071		result = requested;
1072	}
1073	nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
1074out_release:
1075	nfs_direct_req_release(dreq);
1076out:
1077	return result;
1078}
1079
1080/**
1081 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1082 *
1083 */
1084int __init nfs_init_directcache(void)
1085{
1086	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1087						sizeof(struct nfs_direct_req),
1088						0, SLAB_RECLAIM_ACCOUNT,
 
1089						NULL);
1090	if (nfs_direct_cachep == NULL)
1091		return -ENOMEM;
1092
1093	return 0;
1094}
1095
1096/**
1097 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1098 *
1099 */
1100void nfs_destroy_directcache(void)
1101{
1102	kmem_cache_destroy(nfs_direct_cachep);
1103}