Linux Audio

Check our new training course

Loading...
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}
v3.5.6
 
  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
 50#include <linux/nfs_fs.h>
 51#include <linux/nfs_page.h>
 52#include <linux/sunrpc/clnt.h>
 53
 54#include <asm/uaccess.h>
 55#include <linux/atomic.h>
 56
 
 57#include "internal.h"
 58#include "iostat.h"
 59#include "pnfs.h"
 
 
 60
 61#define NFSDBG_FACILITY		NFSDBG_VFS
 62
 63static struct kmem_cache *nfs_direct_cachep;
 64
 65/*
 66 * This represents a set of asynchronous requests that we're waiting on
 67 */
 68struct nfs_direct_req {
 69	struct kref		kref;		/* release manager */
 70
 71	/* I/O parameters */
 72	struct nfs_open_context	*ctx;		/* file open context info */
 73	struct nfs_lock_context *l_ctx;		/* Lock context info */
 74	struct kiocb *		iocb;		/* controlling i/o request */
 75	struct inode *		inode;		/* target file of i/o */
 76
 77	/* completion state */
 78	atomic_t		io_count;	/* i/os we're waiting for */
 79	spinlock_t		lock;		/* protect completion state */
 80	ssize_t			count,		/* bytes actually processed */
 81				error;		/* any reported error */
 82	struct completion	completion;	/* wait for i/o completion */
 83
 84	/* commit state */
 85	struct nfs_mds_commit_info mds_cinfo;	/* Storage for cinfo */
 86	struct pnfs_ds_commit_info ds_cinfo;	/* Storage for cinfo */
 87	struct work_struct	work;
 88	int			flags;
 89#define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
 90#define NFS_ODIRECT_RESCHED_WRITES	(2)	/* write verification failed */
 91	struct nfs_writeverf	verf;		/* unstable write verifier */
 92};
 93
 94static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 95static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 96static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
 97static void nfs_direct_write_schedule_work(struct work_struct *work);
 98
 99static inline void get_dreq(struct nfs_direct_req *dreq)
100{
101	atomic_inc(&dreq->io_count);
102}
103
104static inline int put_dreq(struct nfs_direct_req *dreq)
105{
106	return atomic_dec_and_test(&dreq->io_count);
107}
108
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
109/**
110 * nfs_direct_IO - NFS address space operation for direct I/O
111 * @rw: direction (read or write)
112 * @iocb: target I/O control block
113 * @iov: array of vectors that define I/O buffer
114 * @pos: offset in file to begin the operation
115 * @nr_segs: size of iovec array
116 *
117 * The presence of this routine in the address space ops vector means
118 * the NFS client supports direct I/O.  However, we shunt off direct
119 * read and write requests before the VFS gets them, so this method
120 * should never be called.
121 */
122ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
123{
124	dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
125			iocb->ki_filp->f_path.dentry->d_name.name,
126			(long long) pos, nr_segs);
127
128	return -EINVAL;
 
 
 
 
 
 
129}
130
131static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
132{
133	unsigned int i;
134	for (i = 0; i < npages; i++)
135		page_cache_release(pages[i]);
136}
137
138void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
139			      struct nfs_direct_req *dreq)
140{
141	cinfo->lock = &dreq->lock;
142	cinfo->mds = &dreq->mds_cinfo;
143	cinfo->ds = &dreq->ds_cinfo;
144	cinfo->dreq = dreq;
145	cinfo->completion_ops = &nfs_direct_commit_completion_ops;
146}
147
148static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
149{
150	struct nfs_direct_req *dreq;
151
152	dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
153	if (!dreq)
154		return NULL;
155
156	kref_init(&dreq->kref);
157	kref_get(&dreq->kref);
158	init_completion(&dreq->completion);
159	INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 
160	INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
161	spin_lock_init(&dreq->lock);
162
163	return dreq;
164}
165
166static void nfs_direct_req_free(struct kref *kref)
167{
168	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
169
 
170	if (dreq->l_ctx != NULL)
171		nfs_put_lock_context(dreq->l_ctx);
172	if (dreq->ctx != NULL)
173		put_nfs_open_context(dreq->ctx);
174	kmem_cache_free(nfs_direct_cachep, dreq);
175}
176
177static void nfs_direct_req_release(struct nfs_direct_req *dreq)
178{
179	kref_put(&dreq->kref, nfs_direct_req_free);
180}
181
 
 
 
 
 
 
 
182/*
183 * Collects and returns the final error value/byte-count.
184 */
185static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
186{
187	ssize_t result = -EIOCBQUEUED;
188
189	/* Async requests don't wait here */
190	if (dreq->iocb)
191		goto out;
192
193	result = wait_for_completion_killable(&dreq->completion);
194
 
 
 
 
195	if (!result)
196		result = dreq->error;
197	if (!result)
198		result = dreq->count;
199
200out:
201	return (ssize_t) result;
202}
203
204/*
205 * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
206 * the iocb is still valid here if this is a synchronous request.
207 */
208static void nfs_direct_complete(struct nfs_direct_req *dreq)
209{
 
 
 
 
210	if (dreq->iocb) {
211		long res = (long) dreq->error;
212		if (!res)
213			res = (long) dreq->count;
214		aio_complete(dreq->iocb, res, 0);
 
 
215	}
216	complete_all(&dreq->completion);
 
217
218	nfs_direct_req_release(dreq);
219}
220
221static void nfs_direct_readpage_release(struct nfs_page *req)
222{
223	dprintk("NFS: direct read done (%s/%lld %d@%lld)\n",
224		req->wb_context->dentry->d_inode->i_sb->s_id,
225		(long long)NFS_FILEID(req->wb_context->dentry->d_inode),
226		req->wb_bytes,
227		(long long)req_offset(req));
228	nfs_release_request(req);
229}
230
231static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
232{
233	unsigned long bytes = 0;
234	struct nfs_direct_req *dreq = hdr->dreq;
235
236	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 
 
237		goto out_put;
 
238
239	spin_lock(&dreq->lock);
240	if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
241		dreq->error = hdr->error;
242	else
243		dreq->count += hdr->good_bytes;
244	spin_unlock(&dreq->lock);
245
 
 
246	while (!list_empty(&hdr->pages)) {
247		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
248		struct page *page = req->wb_page;
249
250		if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
251			if (bytes > hdr->good_bytes)
252				zero_user(page, 0, PAGE_SIZE);
253			else if (hdr->good_bytes - bytes < PAGE_SIZE)
254				zero_user_segment(page,
255					hdr->good_bytes & ~PAGE_MASK,
256					PAGE_SIZE);
257		}
258		if (!PageCompound(page)) {
259			if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
260				if (bytes < hdr->good_bytes)
261					set_page_dirty(page);
262			} else
263				set_page_dirty(page);
264		}
265		bytes += req->wb_bytes;
266		nfs_list_remove_request(req);
267		nfs_direct_readpage_release(req);
268	}
269out_put:
270	if (put_dreq(dreq))
271		nfs_direct_complete(dreq);
272	hdr->release(hdr);
273}
274
275static void nfs_read_sync_pgio_error(struct list_head *head)
276{
277	struct nfs_page *req;
278
279	while (!list_empty(head)) {
280		req = nfs_list_entry(head->next);
281		nfs_list_remove_request(req);
282		nfs_release_request(req);
283	}
284}
285
286static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
287{
288	get_dreq(hdr->dreq);
289}
290
291static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
292	.error_cleanup = nfs_read_sync_pgio_error,
293	.init_hdr = nfs_direct_pgio_init,
294	.completion = nfs_direct_read_completion,
295};
296
297/*
298 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
299 * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
300 * bail and stop sending more reads.  Read length accounting is
301 * handled automatically by nfs_direct_read_result().  Otherwise, if
302 * no requests have been sent, just return an error.
303 */
304static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
305						const struct iovec *iov,
306						loff_t pos)
307{
308	struct nfs_direct_req *dreq = desc->pg_dreq;
309	struct nfs_open_context *ctx = dreq->ctx;
310	struct inode *inode = ctx->dentry->d_inode;
311	unsigned long user_addr = (unsigned long)iov->iov_base;
312	size_t count = iov->iov_len;
313	size_t rsize = NFS_SERVER(inode)->rsize;
314	unsigned int pgbase;
315	int result;
316	ssize_t started = 0;
317	struct page **pagevec = NULL;
318	unsigned int npages;
319
320	do {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
321		size_t bytes;
322		int i;
323
324		pgbase = user_addr & ~PAGE_MASK;
325		bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
326
327		result = -ENOMEM;
328		npages = nfs_page_array_len(pgbase, bytes);
329		if (!pagevec)
330			pagevec = kmalloc(npages * sizeof(struct page *),
331					  GFP_KERNEL);
332		if (!pagevec)
333			break;
334		down_read(&current->mm->mmap_sem);
335		result = get_user_pages(current, current->mm, user_addr,
336					npages, 1, 0, pagevec, NULL);
337		up_read(&current->mm->mmap_sem);
338		if (result < 0)
339			break;
340		if ((unsigned)result < npages) {
341			bytes = result * PAGE_SIZE;
342			if (bytes <= pgbase) {
343				nfs_direct_release_pages(pagevec, result);
344				break;
345			}
346			bytes -= pgbase;
347			npages = result;
348		}
349
350		for (i = 0; i < npages; i++) {
351			struct nfs_page *req;
352			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
353			/* XXX do we need to do the eof zeroing found in async_filler? */
354			req = nfs_create_request(dreq->ctx, dreq->inode,
355						 pagevec[i],
356						 pgbase, req_len);
357			if (IS_ERR(req)) {
358				result = PTR_ERR(req);
359				break;
360			}
361			req->wb_index = pos >> PAGE_SHIFT;
362			req->wb_offset = pos & ~PAGE_MASK;
363			if (!nfs_pageio_add_request(desc, req)) {
364				result = desc->pg_error;
365				nfs_release_request(req);
366				break;
367			}
368			pgbase = 0;
369			bytes -= req_len;
370			started += req_len;
371			user_addr += req_len;
372			pos += req_len;
373			count -= req_len;
374		}
375		/* The nfs_page now hold references to these pages */
376		nfs_direct_release_pages(pagevec, npages);
377	} while (count != 0 && result >= 0);
378
379	kfree(pagevec);
380
381	if (started)
382		return started;
383	return result < 0 ? (ssize_t) result : -EFAULT;
384}
385
386static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
387					      const struct iovec *iov,
388					      unsigned long nr_segs,
389					      loff_t pos)
390{
391	struct nfs_pageio_descriptor desc;
392	ssize_t result = -EINVAL;
393	size_t requested_bytes = 0;
394	unsigned long seg;
395
396	nfs_pageio_init_read(&desc, dreq->inode,
397			     &nfs_direct_read_completion_ops);
398	get_dreq(dreq);
399	desc.pg_dreq = dreq;
400
401	for (seg = 0; seg < nr_segs; seg++) {
402		const struct iovec *vec = &iov[seg];
403		result = nfs_direct_read_schedule_segment(&desc, vec, pos);
404		if (result < 0)
405			break;
406		requested_bytes += result;
407		if ((size_t)result < vec->iov_len)
408			break;
409		pos += vec->iov_len;
410	}
411
412	nfs_pageio_complete(&desc);
413
414	/*
415	 * If no bytes were started, return the error, and let the
416	 * generic layer handle the completion.
417	 */
418	if (requested_bytes == 0) {
 
419		nfs_direct_req_release(dreq);
420		return result < 0 ? result : -EIO;
421	}
422
423	if (put_dreq(dreq))
424		nfs_direct_complete(dreq);
425	return 0;
426}
427
428static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
429			       unsigned long nr_segs, loff_t pos)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
430{
431	ssize_t result = -ENOMEM;
432	struct inode *inode = iocb->ki_filp->f_mapping->host;
 
433	struct nfs_direct_req *dreq;
 
 
 
 
 
 
 
434
 
 
 
 
 
 
 
435	dreq = nfs_direct_req_alloc();
436	if (dreq == NULL)
437		goto out;
438
439	dreq->inode = inode;
 
 
440	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
441	dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
442	if (dreq->l_ctx == NULL)
 
 
443		goto out_release;
 
 
444	if (!is_sync_kiocb(iocb))
445		dreq->iocb = iocb;
446
447	result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
448	if (!result)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
449		result = nfs_direct_wait(dreq);
450	NFS_I(inode)->read_io += result;
 
 
 
 
 
 
 
 
451out_release:
452	nfs_direct_req_release(dreq);
453out:
454	return result;
455}
456
457static void nfs_inode_dio_write_done(struct inode *inode)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
458{
459	nfs_zap_mapping(inode, inode->i_mapping);
460	inode_dio_done(inode);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
461}
462
463#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
464static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
465{
466	struct nfs_pageio_descriptor desc;
467	struct nfs_page *req, *tmp;
468	LIST_HEAD(reqs);
469	struct nfs_commit_info cinfo;
470	LIST_HEAD(failed);
471
472	nfs_init_cinfo_from_dreq(&cinfo, dreq);
473	pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
474	spin_lock(cinfo.lock);
475	nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
476	spin_unlock(cinfo.lock);
477
478	dreq->count = 0;
479	get_dreq(dreq);
480
481	nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE,
482			      &nfs_direct_write_completion_ops);
483	desc.pg_dreq = dreq;
484
485	list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 
 
 
486		if (!nfs_pageio_add_request(&desc, req)) {
487			nfs_list_remove_request(req);
488			nfs_list_add_request(req, &failed);
489			spin_lock(cinfo.lock);
490			dreq->flags = 0;
491			dreq->error = -EIO;
492			spin_unlock(cinfo.lock);
 
 
 
 
 
 
493		}
494		nfs_release_request(req);
495	}
496	nfs_pageio_complete(&desc);
497
498	while (!list_empty(&failed)) {
499		req = nfs_list_entry(failed.next);
500		nfs_list_remove_request(req);
501		nfs_unlock_and_release_request(req);
 
 
 
 
 
 
 
 
502	}
503
504	if (put_dreq(dreq))
505		nfs_direct_write_complete(dreq, dreq->inode);
506}
507
508static void nfs_direct_commit_complete(struct nfs_commit_data *data)
509{
 
510	struct nfs_direct_req *dreq = data->dreq;
511	struct nfs_commit_info cinfo;
512	struct nfs_page *req;
513	int status = data->task.tk_status;
514
515	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 
 
516	if (status < 0) {
517		dprintk("NFS: %5u commit failed with error %d.\n",
518			data->task.tk_pid, status);
519		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
520	} else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
521		dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
522		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
523	}
 
 
 
524
525	dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
526	while (!list_empty(&data->pages)) {
527		req = nfs_list_entry(data->pages.next);
528		nfs_list_remove_request(req);
529		if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
530			/* Note the rewrite will go through mds */
531			nfs_mark_request_commit(req, NULL, &cinfo);
 
 
 
 
 
 
 
 
 
 
 
 
 
532		} else
533			nfs_release_request(req);
534		nfs_unlock_and_release_request(req);
535	}
536
537	if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
538		nfs_direct_write_complete(dreq, data->inode);
539}
540
541static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 
542{
543	/* There is no lock to clear */
 
 
 
 
 
 
 
 
544}
545
546static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
547	.completion = nfs_direct_commit_complete,
548	.error_cleanup = nfs_direct_error_cleanup,
549};
550
551static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
552{
553	int res;
554	struct nfs_commit_info cinfo;
555	LIST_HEAD(mds_list);
556
557	nfs_init_cinfo_from_dreq(&cinfo, dreq);
 
558	nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
559	res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
560	if (res < 0) /* res == -ENOMEM */
561		nfs_direct_write_reschedule(dreq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
562}
563
564static void nfs_direct_write_schedule_work(struct work_struct *work)
565{
566	struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
567	int flags = dreq->flags;
568
569	dreq->flags = 0;
570	switch (flags) {
571		case NFS_ODIRECT_DO_COMMIT:
572			nfs_direct_commit_schedule(dreq);
573			break;
574		case NFS_ODIRECT_RESCHED_WRITES:
575			nfs_direct_write_reschedule(dreq);
576			break;
577		default:
578			nfs_inode_dio_write_done(dreq->inode);
 
579			nfs_direct_complete(dreq);
580	}
581}
582
583static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
584{
585	schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
586}
587
588#else
589static void nfs_direct_write_schedule_work(struct work_struct *work)
590{
591}
592
593static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
594{
595	nfs_inode_dio_write_done(inode);
596	nfs_direct_complete(dreq);
597}
598#endif
599
600/*
601 * NB: Return the value of the first error return code.  Subsequent
602 *     errors after the first one are ignored.
603 */
604/*
605 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
606 * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
607 * bail and stop sending more writes.  Write length accounting is
608 * handled automatically by nfs_direct_write_result().  Otherwise, if
609 * no requests have been sent, just return an error.
610 */
611static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
612						 const struct iovec *iov,
613						 loff_t pos)
614{
615	struct nfs_direct_req *dreq = desc->pg_dreq;
616	struct nfs_open_context *ctx = dreq->ctx;
617	struct inode *inode = ctx->dentry->d_inode;
618	unsigned long user_addr = (unsigned long)iov->iov_base;
619	size_t count = iov->iov_len;
620	size_t wsize = NFS_SERVER(inode)->wsize;
621	unsigned int pgbase;
622	int result;
623	ssize_t started = 0;
624	struct page **pagevec = NULL;
625	unsigned int npages;
626
627	do {
628		size_t bytes;
629		int i;
630
631		pgbase = user_addr & ~PAGE_MASK;
632		bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
633
634		result = -ENOMEM;
635		npages = nfs_page_array_len(pgbase, bytes);
636		if (!pagevec)
637			pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
638		if (!pagevec)
639			break;
640
641		down_read(&current->mm->mmap_sem);
642		result = get_user_pages(current, current->mm, user_addr,
643					npages, 0, 0, pagevec, NULL);
644		up_read(&current->mm->mmap_sem);
645		if (result < 0)
646			break;
647
648		if ((unsigned)result < npages) {
649			bytes = result * PAGE_SIZE;
650			if (bytes <= pgbase) {
651				nfs_direct_release_pages(pagevec, result);
652				break;
653			}
654			bytes -= pgbase;
655			npages = result;
656		}
657
658		for (i = 0; i < npages; i++) {
659			struct nfs_page *req;
660			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
661
662			req = nfs_create_request(dreq->ctx, dreq->inode,
663						 pagevec[i],
664						 pgbase, req_len);
665			if (IS_ERR(req)) {
666				result = PTR_ERR(req);
667				break;
668			}
669			nfs_lock_request(req);
670			req->wb_index = pos >> PAGE_SHIFT;
671			req->wb_offset = pos & ~PAGE_MASK;
672			if (!nfs_pageio_add_request(desc, req)) {
673				result = desc->pg_error;
674				nfs_unlock_and_release_request(req);
675				break;
676			}
677			pgbase = 0;
678			bytes -= req_len;
679			started += req_len;
680			user_addr += req_len;
681			pos += req_len;
682			count -= req_len;
683		}
684		/* The nfs_page now hold references to these pages */
685		nfs_direct_release_pages(pagevec, npages);
686	} while (count != 0 && result >= 0);
687
688	kfree(pagevec);
689
690	if (started)
691		return started;
692	return result < 0 ? (ssize_t) result : -EFAULT;
693}
694
695static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
696{
697	struct nfs_direct_req *dreq = hdr->dreq;
698	struct nfs_commit_info cinfo;
699	int bit = -1;
700	struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 
 
701
702	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
703		goto out_put;
704
705	nfs_init_cinfo_from_dreq(&cinfo, dreq);
706
707	spin_lock(&dreq->lock);
 
 
 
 
708
709	if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
710		dreq->flags = 0;
711		dreq->error = hdr->error;
712	}
713	if (dreq->error != 0)
714		bit = NFS_IOHDR_ERROR;
715	else {
716		dreq->count += hdr->good_bytes;
717		if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
718			dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
719			bit = NFS_IOHDR_NEED_RESCHED;
720		} else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
721			if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
722				bit = NFS_IOHDR_NEED_RESCHED;
723			else if (dreq->flags == 0) {
724				memcpy(&dreq->verf, hdr->verf,
725				       sizeof(dreq->verf));
726				bit = NFS_IOHDR_NEED_COMMIT;
727				dreq->flags = NFS_ODIRECT_DO_COMMIT;
728			} else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
729				if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
730					dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
731					bit = NFS_IOHDR_NEED_RESCHED;
732				} else
733					bit = NFS_IOHDR_NEED_COMMIT;
734			}
735		}
736	}
737	spin_unlock(&dreq->lock);
738
 
 
 
 
 
739	while (!list_empty(&hdr->pages)) {
 
740		req = nfs_list_entry(hdr->pages.next);
741		nfs_list_remove_request(req);
742		switch (bit) {
743		case NFS_IOHDR_NEED_RESCHED:
744		case NFS_IOHDR_NEED_COMMIT:
 
 
 
 
745			kref_get(&req->wb_kref);
746			nfs_mark_request_commit(req, hdr->lseg, &cinfo);
747		}
748		nfs_unlock_and_release_request(req);
749	}
750
751out_put:
752	if (put_dreq(dreq))
753		nfs_direct_write_complete(dreq, hdr->inode);
754	hdr->release(hdr);
755}
756
757static void nfs_write_sync_pgio_error(struct list_head *head)
758{
759	struct nfs_page *req;
760
761	while (!list_empty(head)) {
762		req = nfs_list_entry(head->next);
763		nfs_list_remove_request(req);
764		nfs_unlock_and_release_request(req);
765	}
766}
767
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
768static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
769	.error_cleanup = nfs_write_sync_pgio_error,
770	.init_hdr = nfs_direct_pgio_init,
771	.completion = nfs_direct_write_completion,
 
772};
773
 
 
 
 
 
 
 
 
 
 
 
 
774static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
775					       const struct iovec *iov,
776					       unsigned long nr_segs,
777					       loff_t pos)
778{
779	struct nfs_pageio_descriptor desc;
780	struct inode *inode = dreq->inode;
 
781	ssize_t result = 0;
782	size_t requested_bytes = 0;
783	unsigned long seg;
 
 
 
784
785	nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE,
786			      &nfs_direct_write_completion_ops);
787	desc.pg_dreq = dreq;
788	get_dreq(dreq);
789	atomic_inc(&inode->i_dio_count);
790
791	for (seg = 0; seg < nr_segs; seg++) {
792		const struct iovec *vec = &iov[seg];
793		result = nfs_direct_write_schedule_segment(&desc, vec, pos);
 
 
 
 
 
 
794		if (result < 0)
795			break;
796		requested_bytes += result;
797		if ((size_t)result < vec->iov_len)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
798			break;
799		pos += vec->iov_len;
800	}
801	nfs_pageio_complete(&desc);
802	NFS_I(dreq->inode)->write_io += desc.pg_bytes_written;
803
804	/*
805	 * If no bytes were started, return the error, and let the
806	 * generic layer handle the completion.
807	 */
808	if (requested_bytes == 0) {
809		inode_dio_done(inode);
810		nfs_direct_req_release(dreq);
811		return result < 0 ? result : -EIO;
812	}
813
814	if (put_dreq(dreq))
815		nfs_direct_write_complete(dreq, dreq->inode);
816	return 0;
817}
818
819static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
820				unsigned long nr_segs, loff_t pos,
821				size_t count)
822{
823	ssize_t result = -ENOMEM;
824	struct inode *inode = iocb->ki_filp->f_mapping->host;
825	struct nfs_direct_req *dreq;
826
827	dreq = nfs_direct_req_alloc();
828	if (!dreq)
829		goto out;
830
831	dreq->inode = inode;
832	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
833	dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
834	if (dreq->l_ctx == NULL)
835		goto out_release;
836	if (!is_sync_kiocb(iocb))
837		dreq->iocb = iocb;
838
839	result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos);
840	if (!result)
841		result = nfs_direct_wait(dreq);
842out_release:
843	nfs_direct_req_release(dreq);
844out:
845	return result;
846}
847
848/**
849 * nfs_file_direct_read - file direct read operation for NFS files
850 * @iocb: target I/O control block
851 * @iov: vector of user buffers into which to read data
852 * @nr_segs: size of iov vector
853 * @pos: byte offset in file where reading starts
854 *
855 * We use this function for direct reads instead of calling
856 * generic_file_aio_read() in order to avoid gfar's check to see if
857 * the request starts before the end of the file.  For that check
858 * to work, we must generate a GETATTR before each direct read, and
859 * even then there is a window between the GETATTR and the subsequent
860 * READ where the file size could change.  Our preference is simply
861 * to do all reads the application wants, and the server will take
862 * care of managing the end of file boundary.
863 *
864 * This function also eliminates unnecessarily updating the file's
865 * atime locally, as the NFS server sets the file's atime, and this
866 * client must read the updated atime from the server back into its
867 * cache.
868 */
869ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
870				unsigned long nr_segs, loff_t pos)
871{
872	ssize_t retval = -EINVAL;
873	struct file *file = iocb->ki_filp;
874	struct address_space *mapping = file->f_mapping;
875	size_t count;
876
877	count = iov_length(iov, nr_segs);
878	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
879
880	dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
881		file->f_path.dentry->d_parent->d_name.name,
882		file->f_path.dentry->d_name.name,
883		count, (long long) pos);
884
885	retval = 0;
886	if (!count)
887		goto out;
888
889	retval = nfs_sync_mapping(mapping);
890	if (retval)
891		goto out;
892
893	task_io_account_read(count);
894
895	retval = nfs_direct_read(iocb, iov, nr_segs, pos);
896	if (retval > 0)
897		iocb->ki_pos = pos + retval;
898
899out:
900	return retval;
901}
902
903/**
904 * nfs_file_direct_write - file direct write operation for NFS files
905 * @iocb: target I/O control block
906 * @iov: vector of user buffers from which to write data
907 * @nr_segs: size of iov vector
908 * @pos: byte offset in file where writing starts
909 *
910 * We use this function for direct writes instead of calling
911 * generic_file_aio_write() in order to avoid taking the inode
912 * semaphore and updating the i_size.  The NFS server will set
913 * the new i_size and this client must read the updated size
914 * back into its cache.  We let the server do generic write
915 * parameter checking and report problems.
916 *
917 * We eliminate local atime updates, see direct read above.
918 *
919 * We avoid unnecessary page cache invalidations for normal cached
920 * readers of this file.
921 *
922 * Note that O_APPEND is not supported for NFS direct writes, as there
923 * is no atomic O_APPEND write facility in the NFS protocol.
924 */
925ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
926				unsigned long nr_segs, loff_t pos)
927{
928	ssize_t retval = -EINVAL;
 
929	struct file *file = iocb->ki_filp;
930	struct address_space *mapping = file->f_mapping;
931	size_t count;
 
 
 
 
 
 
932
933	count = iov_length(iov, nr_segs);
 
 
 
 
 
 
 
934	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
935
936	dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
937		file->f_path.dentry->d_parent->d_name.name,
938		file->f_path.dentry->d_name.name,
939		count, (long long) pos);
940
941	retval = generic_write_checks(file, &pos, &count, 0);
942	if (retval)
 
943		goto out;
944
945	retval = -EINVAL;
946	if ((ssize_t) count < 0)
947		goto out;
948	retval = 0;
949	if (!count)
950		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
951
952	retval = nfs_sync_mapping(mapping);
953	if (retval)
954		goto out;
 
955
956	task_io_account_write(count);
 
957
958	retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
959	if (retval > 0) {
960		struct inode *inode = mapping->host;
961
962		iocb->ki_pos = pos + retval;
963		spin_lock(&inode->i_lock);
964		if (i_size_read(inode) < iocb->ki_pos)
965			i_size_write(inode, iocb->ki_pos);
966		spin_unlock(&inode->i_lock);
 
 
967	}
 
 
 
968out:
969	return retval;
970}
971
972/**
973 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
974 *
975 */
976int __init nfs_init_directcache(void)
977{
978	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
979						sizeof(struct nfs_direct_req),
980						0, (SLAB_RECLAIM_ACCOUNT|
981							SLAB_MEM_SPREAD),
982						NULL);
983	if (nfs_direct_cachep == NULL)
984		return -ENOMEM;
985
986	return 0;
987}
988
989/**
990 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
991 *
992 */
993void nfs_destroy_directcache(void)
994{
995	kmem_cache_destroy(nfs_direct_cachep);
996}