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_select_verf - select the right verifier
 113  * @dreq - direct request possibly spanning multiple servers
 114  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 115  * @ds_idx - index of data server in data server list, only valid if ds_clp set
 116  *
 117  * returns the correct verifier to use given the role of the server
 118  */
 119 static struct nfs_writeverf *
 120 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 121                        struct nfs_client *ds_clp,
 122                        int ds_idx)
 123 {
 124         struct nfs_writeverf *verfp = &dreq->verf;
 125
 126 #ifdef CONFIG_NFS_V4_1
 127         if (ds_clp) {
 128                 /* pNFS is in use, use the DS verf */
 129                 if (ds_idx >= 0 && ds_idx < dreq->ds_cinfo.nbuckets)
 130                         verfp = &dreq->ds_cinfo.buckets[ds_idx].direct_verf;
 131                 else
 132                         WARN_ON_ONCE(1);
 133         }
 134 #endif
 135         return verfp;
 136 }
 137
 138
 139 /*
 140  * nfs_direct_set_hdr_verf - set the write/commit verifier
 141  * @dreq - direct request possibly spanning multiple servers
 142  * @hdr - pageio header to validate against previously seen verfs
 143  *
 144  * Set the server's (MDS or DS) "seen" verifier
 145  */
 146 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 147                                     struct nfs_pgio_header *hdr)
 148 {
 149         struct nfs_writeverf *verfp;
 150
 151         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp,
 152                                       hdr->ds_idx);
 153         WARN_ON_ONCE(verfp->committed >= 0);
 154         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 155         WARN_ON_ONCE(verfp->committed < 0);
 156 }
 157
 158 /*
 159  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 160  * @dreq - direct request possibly spanning multiple servers
 161  * @hdr - pageio header to validate against previously seen verf
 162  *
 163  * set the server's "seen" verf if not initialized.
 164  * returns result of comparison between @hdr->verf and the "seen"
 165  * verf of the server used by @hdr (DS or MDS)
 166  */
 167 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 168                                           struct nfs_pgio_header *hdr)
 169 {
 170         struct nfs_writeverf *verfp;
 171
 172         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp,
 173                                          hdr->ds_idx);
 174         if (verfp->committed < 0) {
 175                 nfs_direct_set_hdr_verf(dreq, hdr);
 176                 return 0;
 177         }
 178         return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 179 }
 180
 181 /*
 182  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 183  * @dreq - direct request possibly spanning multiple servers
 184  * @data - commit data to validate against previously seen verf
 185  *
 186  * returns result of comparison between @data->verf and the verf of
 187  * the server used by @data (DS or MDS)
 188  */
 189 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 190                                            struct nfs_commit_data *data)
 191 {
 192         struct nfs_writeverf *verfp;
 193
 194         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 195                                          data->ds_commit_index);
 196         WARN_ON_ONCE(verfp->committed < 0);
 197         return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
 198 }
 199
 200 /**
 201  * nfs_direct_IO - NFS address space operation for direct I/O
 202  * @rw: direction (read or write)
 203  * @iocb: target I/O control block
 204  * @iov: array of vectors that define I/O buffer
 205  * @pos: offset in file to begin the operation
 206  * @nr_segs: size of iovec array
 207  *
 208  * The presence of this routine in the address space ops vector means
 209  * the NFS client supports direct I/O. However, for most direct IO, we
 210  * shunt off direct read and write requests before the VFS gets them,
 211  * so this method is only ever called for swap.
 212  */
 213 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
 214 {
 215 #ifndef CONFIG_NFS_SWAP
 216         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 217                         iocb->ki_filp, (long long) pos, iter->nr_segs);
 218
 219         return -EINVAL;
 220 #else
 221         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 222
 223         if (rw == READ || rw == KERNEL_READ)
 224                 return nfs_file_direct_read(iocb, iter, pos,
 225                                 rw == READ ? true : false);
 226         return nfs_file_direct_write(iocb, iter, pos,
 227                                 rw == WRITE ? true : false);
 228 #endif /* CONFIG_NFS_SWAP */
 229 }
 230
 231 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 232 {
 233         unsigned int i;
 234         for (i = 0; i < npages; i++)
 235                 page_cache_release(pages[i]);
 236 }
 237
 238 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 239                               struct nfs_direct_req *dreq)
 240 {
 241         cinfo->lock = &dreq->lock;
 242         cinfo->mds = &dreq->mds_cinfo;
 243         cinfo->ds = &dreq->ds_cinfo;
 244         cinfo->dreq = dreq;
 245         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 246 }
 247
 248 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 249 {
 250         struct nfs_direct_req *dreq;
 251
 252         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 253         if (!dreq)
 254                 return NULL;
 255
 256         kref_init(&dreq->kref);
 257         kref_get(&dreq->kref);
 258         init_completion(&dreq->completion);
 259         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 260         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 261         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 262         spin_lock_init(&dreq->lock);
 263
 264         return dreq;
 265 }
 266
 267 static void nfs_direct_req_free(struct kref *kref)
 268 {
 269         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 270
 271         if (dreq->l_ctx != NULL)
 272                 nfs_put_lock_context(dreq->l_ctx);
 273         if (dreq->ctx != NULL)
 274                 put_nfs_open_context(dreq->ctx);
 275         kmem_cache_free(nfs_direct_cachep, dreq);
 276 }
 277
 278 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 279 {
 280         kref_put(&dreq->kref, nfs_direct_req_free);
 281 }
 282
 283 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 284 {
 285         return dreq->bytes_left;
 286 }
 287 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 288
 289 /*
 290  * Collects and returns the final error value/byte-count.
 291  */
 292 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 293 {
 294         ssize_t result = -EIOCBQUEUED;
 295
 296         /* Async requests don't wait here */
 297         if (dreq->iocb)
 298                 goto out;
 299
 300         result = wait_for_completion_killable(&dreq->completion);
 301
 302         if (!result)
 303                 result = dreq->error;
 304         if (!result)
 305                 result = dreq->count;
 306
 307 out:
 308         return (ssize_t) result;
 309 }
 310
 311 /*
 312  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 313  * the iocb is still valid here if this is a synchronous request.
 314  */
 315 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 316 {
 317         struct inode *inode = dreq->inode;
 318
 319         if (dreq->iocb && write) {
 320                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 321
 322                 spin_lock(&inode->i_lock);
 323                 if (i_size_read(inode) < pos)
 324                         i_size_write(inode, pos);
 325                 spin_unlock(&inode->i_lock);
 326         }
 327
 328         if (write)
 329                 nfs_zap_mapping(inode, inode->i_mapping);
 330
 331         inode_dio_done(inode);
 332
 333         if (dreq->iocb) {
 334                 long res = (long) dreq->error;
 335                 if (!res)
 336                         res = (long) dreq->count;
 337                 aio_complete(dreq->iocb, res, 0);
 338         }
 339
 340         complete_all(&dreq->completion);
 341
 342         nfs_direct_req_release(dreq);
 343 }
 344
 345 static void nfs_direct_readpage_release(struct nfs_page *req)
 346 {
 347         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 348                 req->wb_context->dentry->d_inode->i_sb->s_id,
 349                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 350                 req->wb_bytes,
 351                 (long long)req_offset(req));
 352         nfs_release_request(req);
 353 }
 354
 355 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 356 {
 357         unsigned long bytes = 0;
 358         struct nfs_direct_req *dreq = hdr->dreq;
 359
 360         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 361                 goto out_put;
 362
 363         spin_lock(&dreq->lock);
 364         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 365                 dreq->error = hdr->error;
 366         else
 367                 dreq->count += hdr->good_bytes;
 368         spin_unlock(&dreq->lock);
 369
 370         while (!list_empty(&hdr->pages)) {
 371                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 372                 struct page *page = req->wb_page;
 373
 374                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 375                         set_page_dirty(page);
 376                 bytes += req->wb_bytes;
 377                 nfs_list_remove_request(req);
 378                 nfs_direct_readpage_release(req);
 379         }
 380 out_put:
 381         if (put_dreq(dreq))
 382                 nfs_direct_complete(dreq, false);
 383         hdr->release(hdr);
 384 }
 385
 386 static void nfs_read_sync_pgio_error(struct list_head *head)
 387 {
 388         struct nfs_page *req;
 389
 390         while (!list_empty(head)) {
 391                 req = nfs_list_entry(head->next);
 392                 nfs_list_remove_request(req);
 393                 nfs_release_request(req);
 394         }
 395 }
 396
 397 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 398 {
 399         get_dreq(hdr->dreq);
 400 }
 401
 402 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 403         .error_cleanup = nfs_read_sync_pgio_error,
 404         .init_hdr = nfs_direct_pgio_init,
 405         .completion = nfs_direct_read_completion,
 406 };
 407
 408 /*
 409  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 410  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 411  * bail and stop sending more reads.  Read length accounting is
 412  * handled automatically by nfs_direct_read_result().  Otherwise, if
 413  * no requests have been sent, just return an error.
 414  */
 415
 416 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 417                                               struct iov_iter *iter,
 418                                               loff_t pos)
 419 {
 420         struct nfs_pageio_descriptor desc;
 421         struct inode *inode = dreq->inode;
 422         ssize_t result = -EINVAL;
 423         size_t requested_bytes = 0;
 424         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 425
 426         nfs_pageio_init_read(&desc, dreq->inode, false,
 427                              &nfs_direct_read_completion_ops);
 428         get_dreq(dreq);
 429         desc.pg_dreq = dreq;
 430         atomic_inc(&inode->i_dio_count);
 431
 432         while (iov_iter_count(iter)) {
 433                 struct page **pagevec;
 434                 size_t bytes;
 435                 size_t pgbase;
 436                 unsigned npages, i;
 437
 438                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 439                                                   rsize, &pgbase);
 440                 if (result < 0)
 441                         break;
 442
 443                 bytes = result;
 444                 iov_iter_advance(iter, bytes);
 445                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 446                 for (i = 0; i < npages; i++) {
 447                         struct nfs_page *req;
 448                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 449                         /* XXX do we need to do the eof zeroing found in async_filler? */
 450                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 451                                                  pgbase, req_len);
 452                         if (IS_ERR(req)) {
 453                                 result = PTR_ERR(req);
 454                                 break;
 455                         }
 456                         req->wb_index = pos >> PAGE_SHIFT;
 457                         req->wb_offset = pos & ~PAGE_MASK;
 458                         if (!nfs_pageio_add_request(&desc, req)) {
 459                                 result = desc.pg_error;
 460                                 nfs_release_request(req);
 461                                 break;
 462                         }
 463                         pgbase = 0;
 464                         bytes -= req_len;
 465                         requested_bytes += req_len;
 466                         pos += req_len;
 467                         dreq->bytes_left -= req_len;
 468                 }
 469                 nfs_direct_release_pages(pagevec, npages);
 470                 kvfree(pagevec);
 471                 if (result < 0)
 472                         break;
 473         }
 474
 475         nfs_pageio_complete(&desc);
 476
 477         /*
 478          * If no bytes were started, return the error, and let the
 479          * generic layer handle the completion.
 480          */
 481         if (requested_bytes == 0) {
 482                 inode_dio_done(inode);
 483                 nfs_direct_req_release(dreq);
 484                 return result < 0 ? result : -EIO;
 485         }
 486
 487         if (put_dreq(dreq))
 488                 nfs_direct_complete(dreq, false);
 489         return 0;
 490 }
 491
 492 /**
 493  * nfs_file_direct_read - file direct read operation for NFS files
 494  * @iocb: target I/O control block
 495  * @iter: vector of user buffers into which to read data
 496  * @pos: byte offset in file where reading starts
 497  *
 498  * We use this function for direct reads instead of calling
 499  * generic_file_aio_read() in order to avoid gfar's check to see if
 500  * the request starts before the end of the file.  For that check
 501  * to work, we must generate a GETATTR before each direct read, and
 502  * even then there is a window between the GETATTR and the subsequent
 503  * READ where the file size could change.  Our preference is simply
 504  * to do all reads the application wants, and the server will take
 505  * care of managing the end of file boundary.
 506  *
 507  * This function also eliminates unnecessarily updating the file's
 508  * atime locally, as the NFS server sets the file's atime, and this
 509  * client must read the updated atime from the server back into its
 510  * cache.
 511  */
 512 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 513                                 loff_t pos, bool uio)
 514 {
 515         struct file *file = iocb->ki_filp;
 516         struct address_space *mapping = file->f_mapping;
 517         struct inode *inode = mapping->host;
 518         struct nfs_direct_req *dreq;
 519         struct nfs_lock_context *l_ctx;
 520         ssize_t result = -EINVAL;
 521         size_t count = iov_iter_count(iter);
 522         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 523
 524         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 525                 file, count, (long long) pos);
 526
 527         result = 0;
 528         if (!count)
 529                 goto out;
 530
 531         mutex_lock(&inode->i_mutex);
 532         result = nfs_sync_mapping(mapping);
 533         if (result)
 534                 goto out_unlock;
 535
 536         task_io_account_read(count);
 537
 538         result = -ENOMEM;
 539         dreq = nfs_direct_req_alloc();
 540         if (dreq == NULL)
 541                 goto out_unlock;
 542
 543         dreq->inode = inode;
 544         dreq->bytes_left = count;
 545         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 546         l_ctx = nfs_get_lock_context(dreq->ctx);
 547         if (IS_ERR(l_ctx)) {
 548                 result = PTR_ERR(l_ctx);
 549                 goto out_release;
 550         }
 551         dreq->l_ctx = l_ctx;
 552         if (!is_sync_kiocb(iocb))
 553                 dreq->iocb = iocb;
 554
 555         NFS_I(inode)->read_io += count;
 556         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 557
 558         mutex_unlock(&inode->i_mutex);
 559
 560         if (!result) {
 561                 result = nfs_direct_wait(dreq);
 562                 if (result > 0)
 563                         iocb->ki_pos = pos + result;
 564         }
 565
 566         nfs_direct_req_release(dreq);
 567         return result;
 568
 569 out_release:
 570         nfs_direct_req_release(dreq);
 571 out_unlock:
 572         mutex_unlock(&inode->i_mutex);
 573 out:
 574         return result;
 575 }
 576
 577 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 578 {
 579         struct nfs_pageio_descriptor desc;
 580         struct nfs_page *req, *tmp;
 581         LIST_HEAD(reqs);
 582         struct nfs_commit_info cinfo;
 583         LIST_HEAD(failed);
 584
 585         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 586         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 587         spin_lock(cinfo.lock);
 588         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 589         spin_unlock(cinfo.lock);
 590
 591         dreq->count = 0;
 592         get_dreq(dreq);
 593
 594         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 595                               &nfs_direct_write_completion_ops);
 596         desc.pg_dreq = dreq;
 597
 598         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 599                 if (!nfs_pageio_add_request(&desc, req)) {
 600                         nfs_list_remove_request(req);
 601                         nfs_list_add_request(req, &failed);
 602                         spin_lock(cinfo.lock);
 603                         dreq->flags = 0;
 604                         dreq->error = -EIO;
 605                         spin_unlock(cinfo.lock);
 606                 }
 607                 nfs_release_request(req);
 608         }
 609         nfs_pageio_complete(&desc);
 610
 611         while (!list_empty(&failed)) {
 612                 req = nfs_list_entry(failed.next);
 613                 nfs_list_remove_request(req);
 614                 nfs_unlock_and_release_request(req);
 615         }
 616
 617         if (put_dreq(dreq))
 618                 nfs_direct_write_complete(dreq, dreq->inode);
 619 }
 620
 621 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 622 {
 623         struct nfs_direct_req *dreq = data->dreq;
 624         struct nfs_commit_info cinfo;
 625         struct nfs_page *req;
 626         int status = data->task.tk_status;
 627
 628         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 629         if (status < 0) {
 630                 dprintk("NFS: %5u commit failed with error %d.\n",
 631                         data->task.tk_pid, status);
 632                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 633         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 634                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 635                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 636         }
 637
 638         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 639         while (!list_empty(&data->pages)) {
 640                 req = nfs_list_entry(data->pages.next);
 641                 nfs_list_remove_request(req);
 642                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 643                         /* Note the rewrite will go through mds */
 644                         nfs_mark_request_commit(req, NULL, &cinfo);
 645                 } else
 646                         nfs_release_request(req);
 647                 nfs_unlock_and_release_request(req);
 648         }
 649
 650         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 651                 nfs_direct_write_complete(dreq, data->inode);
 652 }
 653
 654 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 655 {
 656         /* There is no lock to clear */
 657 }
 658
 659 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 660         .completion = nfs_direct_commit_complete,
 661         .error_cleanup = nfs_direct_error_cleanup,
 662 };
 663
 664 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 665 {
 666         int res;
 667         struct nfs_commit_info cinfo;
 668         LIST_HEAD(mds_list);
 669
 670         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 671         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 672         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 673         if (res < 0) /* res == -ENOMEM */
 674                 nfs_direct_write_reschedule(dreq);
 675 }
 676
 677 static void nfs_direct_write_schedule_work(struct work_struct *work)
 678 {
 679         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 680         int flags = dreq->flags;
 681
 682         dreq->flags = 0;
 683         switch (flags) {
 684                 case NFS_ODIRECT_DO_COMMIT:
 685                         nfs_direct_commit_schedule(dreq);
 686                         break;
 687                 case NFS_ODIRECT_RESCHED_WRITES:
 688                         nfs_direct_write_reschedule(dreq);
 689                         break;
 690                 default:
 691                         nfs_direct_complete(dreq, true);
 692         }
 693 }
 694
 695 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 696 {
 697         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 698 }
 699
 700 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 701 {
 702         struct nfs_direct_req *dreq = hdr->dreq;
 703         struct nfs_commit_info cinfo;
 704         bool request_commit = false;
 705         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 706
 707         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 708                 goto out_put;
 709
 710         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 711
 712         spin_lock(&dreq->lock);
 713
 714         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 715                 dreq->flags = 0;
 716                 dreq->error = hdr->error;
 717         }
 718         if (dreq->error == 0) {
 719                 dreq->count += hdr->good_bytes;
 720                 if (nfs_write_need_commit(hdr)) {
 721                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 722                                 request_commit = true;
 723                         else if (dreq->flags == 0) {
 724                                 nfs_direct_set_hdr_verf(dreq, hdr);
 725                                 request_commit = true;
 726                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 727                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 728                                 request_commit = true;
 729                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 730                                         dreq->flags =
 731                                                 NFS_ODIRECT_RESCHED_WRITES;
 732                         }
 733                 }
 734         }
 735         spin_unlock(&dreq->lock);
 736
 737         while (!list_empty(&hdr->pages)) {
 738
 739                 req = nfs_list_entry(hdr->pages.next);
 740                 nfs_list_remove_request(req);
 741                 if (request_commit) {
 742                         kref_get(&req->wb_kref);
 743                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 744                 }
 745                 nfs_unlock_and_release_request(req);
 746         }
 747
 748 out_put:
 749         if (put_dreq(dreq))
 750                 nfs_direct_write_complete(dreq, hdr->inode);
 751         hdr->release(hdr);
 752 }
 753
 754 static void nfs_write_sync_pgio_error(struct list_head *head)
 755 {
 756         struct nfs_page *req;
 757
 758         while (!list_empty(head)) {
 759                 req = nfs_list_entry(head->next);
 760                 nfs_list_remove_request(req);
 761                 nfs_unlock_and_release_request(req);
 762         }
 763 }
 764
 765 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 766         .error_cleanup = nfs_write_sync_pgio_error,
 767         .init_hdr = nfs_direct_pgio_init,
 768         .completion = nfs_direct_write_completion,
 769 };
 770
 771
 772 /*
 773  * NB: Return the value of the first error return code.  Subsequent
 774  *     errors after the first one are ignored.
 775  */
 776 /*
 777  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 778  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 779  * bail and stop sending more writes.  Write length accounting is
 780  * handled automatically by nfs_direct_write_result().  Otherwise, if
 781  * no requests have been sent, just return an error.
 782  */
 783 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 784                                                struct iov_iter *iter,
 785                                                loff_t pos)
 786 {
 787         struct nfs_pageio_descriptor desc;
 788         struct inode *inode = dreq->inode;
 789         ssize_t result = 0;
 790         size_t requested_bytes = 0;
 791         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 792
 793         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 794                               &nfs_direct_write_completion_ops);
 795         desc.pg_dreq = dreq;
 796         get_dreq(dreq);
 797         atomic_inc(&inode->i_dio_count);
 798
 799         NFS_I(inode)->write_io += iov_iter_count(iter);
 800         while (iov_iter_count(iter)) {
 801                 struct page **pagevec;
 802                 size_t bytes;
 803                 size_t pgbase;
 804                 unsigned npages, i;
 805
 806                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 807                                                   wsize, &pgbase);
 808                 if (result < 0)
 809                         break;
 810
 811                 bytes = result;
 812                 iov_iter_advance(iter, bytes);
 813                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 814                 for (i = 0; i < npages; i++) {
 815                         struct nfs_page *req;
 816                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 817
 818                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 819                                                  pgbase, req_len);
 820                         if (IS_ERR(req)) {
 821                                 result = PTR_ERR(req);
 822                                 break;
 823                         }
 824                         nfs_lock_request(req);
 825                         req->wb_index = pos >> PAGE_SHIFT;
 826                         req->wb_offset = pos & ~PAGE_MASK;
 827                         if (!nfs_pageio_add_request(&desc, req)) {
 828                                 result = desc.pg_error;
 829                                 nfs_unlock_and_release_request(req);
 830                                 break;
 831                         }
 832                         pgbase = 0;
 833                         bytes -= req_len;
 834                         requested_bytes += req_len;
 835                         pos += req_len;
 836                         dreq->bytes_left -= req_len;
 837                 }
 838                 nfs_direct_release_pages(pagevec, npages);
 839                 kvfree(pagevec);
 840                 if (result < 0)
 841                         break;
 842         }
 843         nfs_pageio_complete(&desc);
 844
 845         /*
 846          * If no bytes were started, return the error, and let the
 847          * generic layer handle the completion.
 848          */
 849         if (requested_bytes == 0) {
 850                 inode_dio_done(inode);
 851                 nfs_direct_req_release(dreq);
 852                 return result < 0 ? result : -EIO;
 853         }
 854
 855         if (put_dreq(dreq))
 856                 nfs_direct_write_complete(dreq, dreq->inode);
 857         return 0;
 858 }
 859
 860 /**
 861  * nfs_file_direct_write - file direct write operation for NFS files
 862  * @iocb: target I/O control block
 863  * @iter: vector of user buffers from which to write data
 864  * @pos: byte offset in file where writing starts
 865  *
 866  * We use this function for direct writes instead of calling
 867  * generic_file_aio_write() in order to avoid taking the inode
 868  * semaphore and updating the i_size.  The NFS server will set
 869  * the new i_size and this client must read the updated size
 870  * back into its cache.  We let the server do generic write
 871  * parameter checking and report problems.
 872  *
 873  * We eliminate local atime updates, see direct read above.
 874  *
 875  * We avoid unnecessary page cache invalidations for normal cached
 876  * readers of this file.
 877  *
 878  * Note that O_APPEND is not supported for NFS direct writes, as there
 879  * is no atomic O_APPEND write facility in the NFS protocol.
 880  */
 881 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 882                                 loff_t pos, bool uio)
 883 {
 884         ssize_t result = -EINVAL;
 885         struct file *file = iocb->ki_filp;
 886         struct address_space *mapping = file->f_mapping;
 887         struct inode *inode = mapping->host;
 888         struct nfs_direct_req *dreq;
 889         struct nfs_lock_context *l_ctx;
 890         loff_t end;
 891         size_t count = iov_iter_count(iter);
 892         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 893
 894         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 895
 896         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 897                 file, count, (long long) pos);
 898
 899         result = generic_write_checks(file, &pos, &count, 0);
 900         if (result)
 901                 goto out;
 902
 903         result = -EINVAL;
 904         if ((ssize_t) count < 0)
 905                 goto out;
 906         result = 0;
 907         if (!count)
 908                 goto out;
 909
 910         mutex_lock(&inode->i_mutex);
 911
 912         result = nfs_sync_mapping(mapping);
 913         if (result)
 914                 goto out_unlock;
 915
 916         if (mapping->nrpages) {
 917                 result = invalidate_inode_pages2_range(mapping,
 918                                         pos >> PAGE_CACHE_SHIFT, end);
 919                 if (result)
 920                         goto out_unlock;
 921         }
 922
 923         task_io_account_write(count);
 924
 925         result = -ENOMEM;
 926         dreq = nfs_direct_req_alloc();
 927         if (!dreq)
 928                 goto out_unlock;
 929
 930         dreq->inode = inode;
 931         dreq->bytes_left = count;
 932         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 933         l_ctx = nfs_get_lock_context(dreq->ctx);
 934         if (IS_ERR(l_ctx)) {
 935                 result = PTR_ERR(l_ctx);
 936                 goto out_release;
 937         }
 938         dreq->l_ctx = l_ctx;
 939         if (!is_sync_kiocb(iocb))
 940                 dreq->iocb = iocb;
 941
 942         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
 943
 944         if (mapping->nrpages) {
 945                 invalidate_inode_pages2_range(mapping,
 946                                               pos >> PAGE_CACHE_SHIFT, end);
 947         }
 948
 949         mutex_unlock(&inode->i_mutex);
 950
 951         if (!result) {
 952                 result = nfs_direct_wait(dreq);
 953                 if (result > 0) {
 954                         struct inode *inode = mapping->host;
 955
 956                         iocb->ki_pos = pos + result;
 957                         spin_lock(&inode->i_lock);
 958                         if (i_size_read(inode) < iocb->ki_pos)
 959                                 i_size_write(inode, iocb->ki_pos);
 960                         spin_unlock(&inode->i_lock);
 961                 }
 962         }
 963         nfs_direct_req_release(dreq);
 964         return result;
 965
 966 out_release:
 967         nfs_direct_req_release(dreq);
 968 out_unlock:
 969         mutex_unlock(&inode->i_mutex);
 970 out:
 971         return result;
 972 }
 973
 974 /**
 975  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 976  *
 977  */
 978 int __init nfs_init_directcache(void)
 979 {
 980         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 981                                                 sizeof(struct nfs_direct_req),
 982                                                 0, (SLAB_RECLAIM_ACCOUNT|
 983                                                         SLAB_MEM_SPREAD),
 984                                                 NULL);
 985         if (nfs_direct_cachep == NULL)
 986                 return -ENOMEM;
 987
 988         return 0;
 989 }
 990
 991 /**
 992  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 993  *
 994  */
 995 void nfs_destroy_directcache(void)
 996 {
 997         kmem_cache_destroy(nfs_direct_cachep);
 998 }