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