fs/nfs/direct.c

   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 <asm/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
  64 static struct kmem_cache *nfs_direct_cachep;
  65
  66 /*
  67  * This represents a set of asynchronous requests that we're waiting on
  68  */
  69 struct nfs_direct_req {
  70         struct kref             kref;           /* release manager */
  71
  72         /* I/O parameters */
  73         struct nfs_open_context *ctx;           /* file open context info */
  74         struct nfs_lock_context *l_ctx;         /* Lock context info */
  75         struct kiocb *          iocb;           /* controlling i/o request */
  76         struct inode *          inode;          /* target file of i/o */
  77
  78         /* completion state */
  79         atomic_t                io_count;       /* i/os we're waiting for */
  80         spinlock_t              lock;           /* protect completion state */
  81         ssize_t                 count,          /* bytes actually processed */
  82                                 bytes_left,     /* bytes left to be sent */
  83                                 error;          /* any reported error */
  84         struct completion       completion;     /* wait for i/o completion */
  85
  86         /* commit state */
  87         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  88         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  89         struct work_struct      work;
  90         int                     flags;
  91 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  92 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  93         struct nfs_writeverf    verf;           /* unstable write verifier */
  94 };
  95
  96 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  97 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  98 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  99 static void nfs_direct_write_schedule_work(struct work_struct *work);
 100
 101 static inline void get_dreq(struct nfs_direct_req *dreq)
 102 {
 103         atomic_inc(&dreq->io_count);
 104 }
 105
 106 static inline int put_dreq(struct nfs_direct_req *dreq)
 107 {
 108         return atomic_dec_and_test(&dreq->io_count);
 109 }
 110
 111 /**
 112  * nfs_direct_IO - NFS address space operation for direct I/O
 113  * @rw: direction (read or write)
 114  * @iocb: target I/O control block
 115  * @iov: array of vectors that define I/O buffer
 116  * @pos: offset in file to begin the operation
 117  * @nr_segs: size of iovec array
 118  *
 119  * The presence of this routine in the address space ops vector means
 120  * the NFS client supports direct I/O. However, for most direct IO, we
 121  * shunt off direct read and write requests before the VFS gets them,
 122  * so this method is only ever called for swap.
 123  */
 124 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 125 {
 126 #ifndef CONFIG_NFS_SWAP
 127         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 128                         iocb->ki_filp, (long long) pos, nr_segs);
 129
 130         return -EINVAL;
 131 #else
 132         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 133
 134         if (rw == READ || rw == KERNEL_READ)
 135                 return nfs_file_direct_read(iocb, iov, nr_segs, pos,
 136                                 rw == READ ? true : false);
 137         return nfs_file_direct_write(iocb, iov, nr_segs, pos,
 138                                 rw == WRITE ? true : false);
 139 #endif /* CONFIG_NFS_SWAP */
 140 }
 141
 142 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 143 {
 144         unsigned int i;
 145         for (i = 0; i < npages; i++)
 146                 page_cache_release(pages[i]);
 147 }
 148
 149 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 150                               struct nfs_direct_req *dreq)
 151 {
 152         cinfo->lock = &dreq->lock;
 153         cinfo->mds = &dreq->mds_cinfo;
 154         cinfo->ds = &dreq->ds_cinfo;
 155         cinfo->dreq = dreq;
 156         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 157 }
 158
 159 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 160 {
 161         struct nfs_direct_req *dreq;
 162
 163         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 164         if (!dreq)
 165                 return NULL;
 166
 167         kref_init(&dreq->kref);
 168         kref_get(&dreq->kref);
 169         init_completion(&dreq->completion);
 170         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 171         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 172         spin_lock_init(&dreq->lock);
 173
 174         return dreq;
 175 }
 176
 177 static void nfs_direct_req_free(struct kref *kref)
 178 {
 179         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 180
 181         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 182         if (dreq->l_ctx != NULL)
 183                 nfs_put_lock_context(dreq->l_ctx);
 184         if (dreq->ctx != NULL)
 185                 put_nfs_open_context(dreq->ctx);
 186         kmem_cache_free(nfs_direct_cachep, dreq);
 187 }
 188
 189 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 190 {
 191         kref_put(&dreq->kref, nfs_direct_req_free);
 192 }
 193
 194 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 195 {
 196         return dreq->bytes_left;
 197 }
 198 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 199
 200 /*
 201  * Collects and returns the final error value/byte-count.
 202  */
 203 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 204 {
 205         ssize_t result = -EIOCBQUEUED;
 206
 207         /* Async requests don't wait here */
 208         if (dreq->iocb)
 209                 goto out;
 210
 211         result = wait_for_completion_killable(&dreq->completion);
 212
 213         if (!result)
 214                 result = dreq->error;
 215         if (!result)
 216                 result = dreq->count;
 217
 218 out:
 219         return (ssize_t) result;
 220 }
 221
 222 /*
 223  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 224  * the iocb is still valid here if this is a synchronous request.
 225  */
 226 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 227 {
 228         struct inode *inode = dreq->inode;
 229
 230         if (dreq->iocb && write) {
 231                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 232
 233                 spin_lock(&inode->i_lock);
 234                 if (i_size_read(inode) < pos)
 235                         i_size_write(inode, pos);
 236                 spin_unlock(&inode->i_lock);
 237         }
 238
 239         if (write)
 240                 nfs_zap_mapping(inode, inode->i_mapping);
 241
 242         inode_dio_done(inode);
 243
 244         if (dreq->iocb) {
 245                 long res = (long) dreq->error;
 246                 if (!res)
 247                         res = (long) dreq->count;
 248                 aio_complete(dreq->iocb, res, 0);
 249         }
 250
 251         complete_all(&dreq->completion);
 252
 253         nfs_direct_req_release(dreq);
 254 }
 255
 256 static void nfs_direct_readpage_release(struct nfs_page *req)
 257 {
 258         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 259                 req->wb_context->dentry->d_inode->i_sb->s_id,
 260                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 261                 req->wb_bytes,
 262                 (long long)req_offset(req));
 263         nfs_release_request(req);
 264 }
 265
 266 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 267 {
 268         unsigned long bytes = 0;
 269         struct nfs_direct_req *dreq = hdr->dreq;
 270
 271         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 272                 goto out_put;
 273
 274         spin_lock(&dreq->lock);
 275         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 276                 dreq->error = hdr->error;
 277         else
 278                 dreq->count += hdr->good_bytes;
 279         spin_unlock(&dreq->lock);
 280
 281         while (!list_empty(&hdr->pages)) {
 282                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 283                 struct page *page = req->wb_page;
 284
 285                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 286                         set_page_dirty(page);
 287                 bytes += req->wb_bytes;
 288                 nfs_list_remove_request(req);
 289                 nfs_direct_readpage_release(req);
 290         }
 291 out_put:
 292         if (put_dreq(dreq))
 293                 nfs_direct_complete(dreq, false);
 294         hdr->release(hdr);
 295 }
 296
 297 static void nfs_read_sync_pgio_error(struct list_head *head)
 298 {
 299         struct nfs_page *req;
 300
 301         while (!list_empty(head)) {
 302                 req = nfs_list_entry(head->next);
 303                 nfs_list_remove_request(req);
 304                 nfs_release_request(req);
 305         }
 306 }
 307
 308 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 309 {
 310         get_dreq(hdr->dreq);
 311 }
 312
 313 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 314         .error_cleanup = nfs_read_sync_pgio_error,
 315         .init_hdr = nfs_direct_pgio_init,
 316         .completion = nfs_direct_read_completion,
 317 };
 318
 319 /*
 320  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 321  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 322  * bail and stop sending more reads.  Read length accounting is
 323  * handled automatically by nfs_direct_read_result().  Otherwise, if
 324  * no requests have been sent, just return an error.
 325  */
 326 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
 327                                                 const struct iovec *iov,
 328                                                 loff_t pos, bool uio)
 329 {
 330         struct nfs_direct_req *dreq = desc->pg_dreq;
 331         struct nfs_open_context *ctx = dreq->ctx;
 332         struct inode *inode = ctx->dentry->d_inode;
 333         unsigned long user_addr = (unsigned long)iov->iov_base;
 334         size_t count = iov->iov_len;
 335         size_t rsize = NFS_SERVER(inode)->rsize;
 336         unsigned int pgbase;
 337         int result;
 338         ssize_t started = 0;
 339         struct page **pagevec = NULL;
 340         unsigned int npages;
 341
 342         do {
 343                 size_t bytes;
 344                 int i;
 345
 346                 pgbase = user_addr & ~PAGE_MASK;
 347                 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
 348
 349                 result = -ENOMEM;
 350                 npages = nfs_page_array_len(pgbase, bytes);
 351                 if (!pagevec)
 352                         pagevec = kmalloc(npages * sizeof(struct page *),
 353                                           GFP_KERNEL);
 354                 if (!pagevec)
 355                         break;
 356                 if (uio) {
 357                         down_read(&current->mm->mmap_sem);
 358                         result = get_user_pages(current, current->mm, user_addr,
 359                                         npages, 1, 0, pagevec, NULL);
 360                         up_read(&current->mm->mmap_sem);
 361                         if (result < 0)
 362                                 break;
 363                 } else {
 364                         WARN_ON(npages != 1);
 365                         result = get_kernel_page(user_addr, 1, pagevec);
 366                         if (WARN_ON(result != 1))
 367                                 break;
 368                 }
 369
 370                 if ((unsigned)result < npages) {
 371                         bytes = result * PAGE_SIZE;
 372                         if (bytes <= pgbase) {
 373                                 nfs_direct_release_pages(pagevec, result);
 374                                 break;
 375                         }
 376                         bytes -= pgbase;
 377                         npages = result;
 378                 }
 379
 380                 for (i = 0; i < npages; i++) {
 381                         struct nfs_page *req;
 382                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 383                         /* XXX do we need to do the eof zeroing found in async_filler? */
 384                         req = nfs_create_request(dreq->ctx, dreq->inode,
 385                                                  pagevec[i],
 386                                                  pgbase, req_len);
 387                         if (IS_ERR(req)) {
 388                                 result = PTR_ERR(req);
 389                                 break;
 390                         }
 391                         req->wb_index = pos >> PAGE_SHIFT;
 392                         req->wb_offset = pos & ~PAGE_MASK;
 393                         if (!nfs_pageio_add_request(desc, req)) {
 394                                 result = desc->pg_error;
 395                                 nfs_release_request(req);
 396                                 break;
 397                         }
 398                         pgbase = 0;
 399                         bytes -= req_len;
 400                         started += req_len;
 401                         user_addr += req_len;
 402                         pos += req_len;
 403                         count -= req_len;
 404                         dreq->bytes_left -= req_len;
 405                 }
 406                 /* The nfs_page now hold references to these pages */
 407                 nfs_direct_release_pages(pagevec, npages);
 408         } while (count != 0 && result >= 0);
 409
 410         kfree(pagevec);
 411
 412         if (started)
 413                 return started;
 414         return result < 0 ? (ssize_t) result : -EFAULT;
 415 }
 416
 417 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 418                                               const struct iovec *iov,
 419                                               unsigned long nr_segs,
 420                                               loff_t pos, bool uio)
 421 {
 422         struct nfs_pageio_descriptor desc;
 423         struct inode *inode = dreq->inode;
 424         ssize_t result = -EINVAL;
 425         size_t requested_bytes = 0;
 426         unsigned long seg;
 427
 428         NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
 429                              &nfs_direct_read_completion_ops);
 430         get_dreq(dreq);
 431         desc.pg_dreq = dreq;
 432         atomic_inc(&inode->i_dio_count);
 433
 434         for (seg = 0; seg < nr_segs; seg++) {
 435                 const struct iovec *vec = &iov[seg];
 436                 result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio);
 437                 if (result < 0)
 438                         break;
 439                 requested_bytes += result;
 440                 if ((size_t)result < vec->iov_len)
 441                         break;
 442                 pos += vec->iov_len;
 443         }
 444
 445         nfs_pageio_complete(&desc);
 446
 447         /*
 448          * If no bytes were started, return the error, and let the
 449          * generic layer handle the completion.
 450          */
 451         if (requested_bytes == 0) {
 452                 inode_dio_done(inode);
 453                 nfs_direct_req_release(dreq);
 454                 return result < 0 ? result : -EIO;
 455         }
 456
 457         if (put_dreq(dreq))
 458                 nfs_direct_complete(dreq, false);
 459         return 0;
 460 }
 461
 462 /**
 463  * nfs_file_direct_read - file direct read operation for NFS files
 464  * @iocb: target I/O control block
 465  * @iov: vector of user buffers into which to read data
 466  * @nr_segs: size of iov vector
 467  * @pos: byte offset in file where reading starts
 468  *
 469  * We use this function for direct reads instead of calling
 470  * generic_file_aio_read() in order to avoid gfar's check to see if
 471  * the request starts before the end of the file.  For that check
 472  * to work, we must generate a GETATTR before each direct read, and
 473  * even then there is a window between the GETATTR and the subsequent
 474  * READ where the file size could change.  Our preference is simply
 475  * to do all reads the application wants, and the server will take
 476  * care of managing the end of file boundary.
 477  *
 478  * This function also eliminates unnecessarily updating the file's
 479  * atime locally, as the NFS server sets the file's atime, and this
 480  * client must read the updated atime from the server back into its
 481  * cache.
 482  */
 483 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 484                                 unsigned long nr_segs, loff_t pos, bool uio)
 485 {
 486         struct file *file = iocb->ki_filp;
 487         struct address_space *mapping = file->f_mapping;
 488         struct inode *inode = mapping->host;
 489         struct nfs_direct_req *dreq;
 490         struct nfs_lock_context *l_ctx;
 491         ssize_t result = -EINVAL;
 492         size_t count;
 493
 494         count = iov_length(iov, nr_segs);
 495         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 496
 497         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 498                 file, count, (long long) pos);
 499
 500         result = 0;
 501         if (!count)
 502                 goto out;
 503
 504         mutex_lock(&inode->i_mutex);
 505         result = nfs_sync_mapping(mapping);
 506         if (result)
 507                 goto out_unlock;
 508
 509         task_io_account_read(count);
 510
 511         result = -ENOMEM;
 512         dreq = nfs_direct_req_alloc();
 513         if (dreq == NULL)
 514                 goto out_unlock;
 515
 516         dreq->inode = inode;
 517         dreq->bytes_left = iov_length(iov, nr_segs);
 518         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 519         l_ctx = nfs_get_lock_context(dreq->ctx);
 520         if (IS_ERR(l_ctx)) {
 521                 result = PTR_ERR(l_ctx);
 522                 goto out_release;
 523         }
 524         dreq->l_ctx = l_ctx;
 525         if (!is_sync_kiocb(iocb))
 526                 dreq->iocb = iocb;
 527
 528         NFS_I(inode)->read_io += iov_length(iov, nr_segs);
 529         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio);
 530
 531         mutex_unlock(&inode->i_mutex);
 532
 533         if (!result) {
 534                 result = nfs_direct_wait(dreq);
 535                 if (result > 0)
 536                         iocb->ki_pos = pos + result;
 537         }
 538
 539         nfs_direct_req_release(dreq);
 540         return result;
 541
 542 out_release:
 543         nfs_direct_req_release(dreq);
 544 out_unlock:
 545         mutex_unlock(&inode->i_mutex);
 546 out:
 547         return result;
 548 }
 549
 550 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 551 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 552 {
 553         struct nfs_pageio_descriptor desc;
 554         struct nfs_page *req, *tmp;
 555         LIST_HEAD(reqs);
 556         struct nfs_commit_info cinfo;
 557         LIST_HEAD(failed);
 558
 559         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 560         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 561         spin_lock(cinfo.lock);
 562         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 563         spin_unlock(cinfo.lock);
 564
 565         dreq->count = 0;
 566         get_dreq(dreq);
 567
 568         NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
 569                               &nfs_direct_write_completion_ops);
 570         desc.pg_dreq = dreq;
 571
 572         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 573                 if (!nfs_pageio_add_request(&desc, req)) {
 574                         nfs_list_remove_request(req);
 575                         nfs_list_add_request(req, &failed);
 576                         spin_lock(cinfo.lock);
 577                         dreq->flags = 0;
 578                         dreq->error = -EIO;
 579                         spin_unlock(cinfo.lock);
 580                 }
 581                 nfs_release_request(req);
 582         }
 583         nfs_pageio_complete(&desc);
 584
 585         while (!list_empty(&failed)) {
 586                 req = nfs_list_entry(failed.next);
 587                 nfs_list_remove_request(req);
 588                 nfs_unlock_and_release_request(req);
 589         }
 590
 591         if (put_dreq(dreq))
 592                 nfs_direct_write_complete(dreq, dreq->inode);
 593 }
 594
 595 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 596 {
 597         struct nfs_direct_req *dreq = data->dreq;
 598         struct nfs_commit_info cinfo;
 599         struct nfs_page *req;
 600         int status = data->task.tk_status;
 601
 602         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 603         if (status < 0) {
 604                 dprintk("NFS: %5u commit failed with error %d.\n",
 605                         data->task.tk_pid, status);
 606                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 607         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 608                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 609                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 610         }
 611
 612         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 613         while (!list_empty(&data->pages)) {
 614                 req = nfs_list_entry(data->pages.next);
 615                 nfs_list_remove_request(req);
 616                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 617                         /* Note the rewrite will go through mds */
 618                         nfs_mark_request_commit(req, NULL, &cinfo);
 619                 } else
 620                         nfs_release_request(req);
 621                 nfs_unlock_and_release_request(req);
 622         }
 623
 624         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 625                 nfs_direct_write_complete(dreq, data->inode);
 626 }
 627
 628 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 629 {
 630         /* There is no lock to clear */
 631 }
 632
 633 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 634         .completion = nfs_direct_commit_complete,
 635         .error_cleanup = nfs_direct_error_cleanup,
 636 };
 637
 638 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 639 {
 640         int res;
 641         struct nfs_commit_info cinfo;
 642         LIST_HEAD(mds_list);
 643
 644         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 645         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 646         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 647         if (res < 0) /* res == -ENOMEM */
 648                 nfs_direct_write_reschedule(dreq);
 649 }
 650
 651 static void nfs_direct_write_schedule_work(struct work_struct *work)
 652 {
 653         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 654         int flags = dreq->flags;
 655
 656         dreq->flags = 0;
 657         switch (flags) {
 658                 case NFS_ODIRECT_DO_COMMIT:
 659                         nfs_direct_commit_schedule(dreq);
 660                         break;
 661                 case NFS_ODIRECT_RESCHED_WRITES:
 662                         nfs_direct_write_reschedule(dreq);
 663                         break;
 664                 default:
 665                         nfs_direct_complete(dreq, true);
 666         }
 667 }
 668
 669 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 670 {
 671         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 672 }
 673
 674 #else
 675 static void nfs_direct_write_schedule_work(struct work_struct *work)
 676 {
 677 }
 678
 679 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 680 {
 681         nfs_direct_complete(dreq, true);
 682 }
 683 #endif
 684
 685 /*
 686  * NB: Return the value of the first error return code.  Subsequent
 687  *     errors after the first one are ignored.
 688  */
 689 /*
 690  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 691  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 692  * bail and stop sending more writes.  Write length accounting is
 693  * handled automatically by nfs_direct_write_result().  Otherwise, if
 694  * no requests have been sent, just return an error.
 695  */
 696 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
 697                                                  const struct iovec *iov,
 698                                                  loff_t pos, bool uio)
 699 {
 700         struct nfs_direct_req *dreq = desc->pg_dreq;
 701         struct nfs_open_context *ctx = dreq->ctx;
 702         struct inode *inode = ctx->dentry->d_inode;
 703         unsigned long user_addr = (unsigned long)iov->iov_base;
 704         size_t count = iov->iov_len;
 705         size_t wsize = NFS_SERVER(inode)->wsize;
 706         unsigned int pgbase;
 707         int result;
 708         ssize_t started = 0;
 709         struct page **pagevec = NULL;
 710         unsigned int npages;
 711
 712         do {
 713                 size_t bytes;
 714                 int i;
 715
 716                 pgbase = user_addr & ~PAGE_MASK;
 717                 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
 718
 719                 result = -ENOMEM;
 720                 npages = nfs_page_array_len(pgbase, bytes);
 721                 if (!pagevec)
 722                         pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
 723                 if (!pagevec)
 724                         break;
 725
 726                 if (uio) {
 727                         down_read(&current->mm->mmap_sem);
 728                         result = get_user_pages(current, current->mm, user_addr,
 729                                                 npages, 0, 0, pagevec, NULL);
 730                         up_read(&current->mm->mmap_sem);
 731                         if (result < 0)
 732                                 break;
 733                 } else {
 734                         WARN_ON(npages != 1);
 735                         result = get_kernel_page(user_addr, 0, pagevec);
 736                         if (WARN_ON(result != 1))
 737                                 break;
 738                 }
 739
 740                 if ((unsigned)result < npages) {
 741                         bytes = result * PAGE_SIZE;
 742                         if (bytes <= pgbase) {
 743                                 nfs_direct_release_pages(pagevec, result);
 744                                 break;
 745                         }
 746                         bytes -= pgbase;
 747                         npages = result;
 748                 }
 749
 750                 for (i = 0; i < npages; i++) {
 751                         struct nfs_page *req;
 752                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 753
 754                         req = nfs_create_request(dreq->ctx, dreq->inode,
 755                                                  pagevec[i],
 756                                                  pgbase, req_len);
 757                         if (IS_ERR(req)) {
 758                                 result = PTR_ERR(req);
 759                                 break;
 760                         }
 761                         nfs_lock_request(req);
 762                         req->wb_index = pos >> PAGE_SHIFT;
 763                         req->wb_offset = pos & ~PAGE_MASK;
 764                         if (!nfs_pageio_add_request(desc, req)) {
 765                                 result = desc->pg_error;
 766                                 nfs_unlock_and_release_request(req);
 767                                 break;
 768                         }
 769                         pgbase = 0;
 770                         bytes -= req_len;
 771                         started += req_len;
 772                         user_addr += req_len;
 773                         pos += req_len;
 774                         count -= req_len;
 775                         dreq->bytes_left -= req_len;
 776                 }
 777                 /* The nfs_page now hold references to these pages */
 778                 nfs_direct_release_pages(pagevec, npages);
 779         } while (count != 0 && result >= 0);
 780
 781         kfree(pagevec);
 782
 783         if (started)
 784                 return started;
 785         return result < 0 ? (ssize_t) result : -EFAULT;
 786 }
 787
 788 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 789 {
 790         struct nfs_direct_req *dreq = hdr->dreq;
 791         struct nfs_commit_info cinfo;
 792         int bit = -1;
 793         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 794
 795         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 796                 goto out_put;
 797
 798         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 799
 800         spin_lock(&dreq->lock);
 801
 802         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 803                 dreq->flags = 0;
 804                 dreq->error = hdr->error;
 805         }
 806         if (dreq->error != 0)
 807                 bit = NFS_IOHDR_ERROR;
 808         else {
 809                 dreq->count += hdr->good_bytes;
 810                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 811                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 812                         bit = NFS_IOHDR_NEED_RESCHED;
 813                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 814                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 815                                 bit = NFS_IOHDR_NEED_RESCHED;
 816                         else if (dreq->flags == 0) {
 817                                 memcpy(&dreq->verf, hdr->verf,
 818                                        sizeof(dreq->verf));
 819                                 bit = NFS_IOHDR_NEED_COMMIT;
 820                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 821                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 822                                 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
 823                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 824                                         bit = NFS_IOHDR_NEED_RESCHED;
 825                                 } else
 826                                         bit = NFS_IOHDR_NEED_COMMIT;
 827                         }
 828                 }
 829         }
 830         spin_unlock(&dreq->lock);
 831
 832         while (!list_empty(&hdr->pages)) {
 833                 req = nfs_list_entry(hdr->pages.next);
 834                 nfs_list_remove_request(req);
 835                 switch (bit) {
 836                 case NFS_IOHDR_NEED_RESCHED:
 837                 case NFS_IOHDR_NEED_COMMIT:
 838                         kref_get(&req->wb_kref);
 839                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 840                 }
 841                 nfs_unlock_and_release_request(req);
 842         }
 843
 844 out_put:
 845         if (put_dreq(dreq))
 846                 nfs_direct_write_complete(dreq, hdr->inode);
 847         hdr->release(hdr);
 848 }
 849
 850 static void nfs_write_sync_pgio_error(struct list_head *head)
 851 {
 852         struct nfs_page *req;
 853
 854         while (!list_empty(head)) {
 855                 req = nfs_list_entry(head->next);
 856                 nfs_list_remove_request(req);
 857                 nfs_unlock_and_release_request(req);
 858         }
 859 }
 860
 861 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 862         .error_cleanup = nfs_write_sync_pgio_error,
 863         .init_hdr = nfs_direct_pgio_init,
 864         .completion = nfs_direct_write_completion,
 865 };
 866
 867 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 868                                                const struct iovec *iov,
 869                                                unsigned long nr_segs,
 870                                                loff_t pos, bool uio)
 871 {
 872         struct nfs_pageio_descriptor desc;
 873         struct inode *inode = dreq->inode;
 874         ssize_t result = 0;
 875         size_t requested_bytes = 0;
 876         unsigned long seg;
 877
 878         NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
 879                               &nfs_direct_write_completion_ops);
 880         desc.pg_dreq = dreq;
 881         get_dreq(dreq);
 882         atomic_inc(&inode->i_dio_count);
 883
 884         NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs);
 885         for (seg = 0; seg < nr_segs; seg++) {
 886                 const struct iovec *vec = &iov[seg];
 887                 result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio);
 888                 if (result < 0)
 889                         break;
 890                 requested_bytes += result;
 891                 if ((size_t)result < vec->iov_len)
 892                         break;
 893                 pos += vec->iov_len;
 894         }
 895         nfs_pageio_complete(&desc);
 896
 897         /*
 898          * If no bytes were started, return the error, and let the
 899          * generic layer handle the completion.
 900          */
 901         if (requested_bytes == 0) {
 902                 inode_dio_done(inode);
 903                 nfs_direct_req_release(dreq);
 904                 return result < 0 ? result : -EIO;
 905         }
 906
 907         if (put_dreq(dreq))
 908                 nfs_direct_write_complete(dreq, dreq->inode);
 909         return 0;
 910 }
 911
 912 /**
 913  * nfs_file_direct_write - file direct write operation for NFS files
 914  * @iocb: target I/O control block
 915  * @iov: vector of user buffers from which to write data
 916  * @nr_segs: size of iov vector
 917  * @pos: byte offset in file where writing starts
 918  *
 919  * We use this function for direct writes instead of calling
 920  * generic_file_aio_write() in order to avoid taking the inode
 921  * semaphore and updating the i_size.  The NFS server will set
 922  * the new i_size and this client must read the updated size
 923  * back into its cache.  We let the server do generic write
 924  * parameter checking and report problems.
 925  *
 926  * We eliminate local atime updates, see direct read above.
 927  *
 928  * We avoid unnecessary page cache invalidations for normal cached
 929  * readers of this file.
 930  *
 931  * Note that O_APPEND is not supported for NFS direct writes, as there
 932  * is no atomic O_APPEND write facility in the NFS protocol.
 933  */
 934 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 935                                 unsigned long nr_segs, loff_t pos, bool uio)
 936 {
 937         ssize_t result = -EINVAL;
 938         struct file *file = iocb->ki_filp;
 939         struct address_space *mapping = file->f_mapping;
 940         struct inode *inode = mapping->host;
 941         struct nfs_direct_req *dreq;
 942         struct nfs_lock_context *l_ctx;
 943         loff_t end;
 944         size_t count;
 945
 946         count = iov_length(iov, nr_segs);
 947         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 948
 949         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 950
 951         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 952                 file, count, (long long) pos);
 953
 954         result = generic_write_checks(file, &pos, &count, 0);
 955         if (result)
 956                 goto out;
 957
 958         result = -EINVAL;
 959         if ((ssize_t) count < 0)
 960                 goto out;
 961         result = 0;
 962         if (!count)
 963                 goto out;
 964
 965         mutex_lock(&inode->i_mutex);
 966
 967         result = nfs_sync_mapping(mapping);
 968         if (result)
 969                 goto out_unlock;
 970
 971         if (mapping->nrpages) {
 972                 result = invalidate_inode_pages2_range(mapping,
 973                                         pos >> PAGE_CACHE_SHIFT, end);
 974                 if (result)
 975                         goto out_unlock;
 976         }
 977
 978         task_io_account_write(count);
 979
 980         result = -ENOMEM;
 981         dreq = nfs_direct_req_alloc();
 982         if (!dreq)
 983                 goto out_unlock;
 984
 985         dreq->inode = inode;
 986         dreq->bytes_left = count;
 987         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 988         l_ctx = nfs_get_lock_context(dreq->ctx);
 989         if (IS_ERR(l_ctx)) {
 990                 result = PTR_ERR(l_ctx);
 991                 goto out_release;
 992         }
 993         dreq->l_ctx = l_ctx;
 994         if (!is_sync_kiocb(iocb))
 995                 dreq->iocb = iocb;
 996
 997         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio);
 998
 999         if (mapping->nrpages) {
1000                 invalidate_inode_pages2_range(mapping,
1001                                               pos >> PAGE_CACHE_SHIFT, end);
1002         }
1003
1004         mutex_unlock(&inode->i_mutex);
1005
1006         if (!result) {
1007                 result = nfs_direct_wait(dreq);
1008                 if (result > 0) {
1009                         struct inode *inode = mapping->host;
1010
1011                         iocb->ki_pos = pos + result;
1012                         spin_lock(&inode->i_lock);
1013                         if (i_size_read(inode) < iocb->ki_pos)
1014                                 i_size_write(inode, iocb->ki_pos);
1015                         spin_unlock(&inode->i_lock);
1016                 }
1017         }
1018         nfs_direct_req_release(dreq);
1019         return result;
1020
1021 out_release:
1022         nfs_direct_req_release(dreq);
1023 out_unlock:
1024         mutex_unlock(&inode->i_mutex);
1025 out:
1026         return result;
1027 }
1028
1029 /**
1030  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1031  *
1032  */
1033 int __init nfs_init_directcache(void)
1034 {
1035         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1036                                                 sizeof(struct nfs_direct_req),
1037                                                 0, (SLAB_RECLAIM_ACCOUNT|
1038                                                         SLAB_MEM_SPREAD),
1039                                                 NULL);
1040         if (nfs_direct_cachep == NULL)
1041                 return -ENOMEM;
1042
1043         return 0;
1044 }
1045
1046 /**
1047  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1048  *
1049  */
1050 void nfs_destroy_directcache(void)
1051 {
1052         kmem_cache_destroy(nfs_direct_cachep);
1053 }