fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <[email protected]>
   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)
 224                 return nfs_file_direct_read(iocb, iter, pos);
 225         return nfs_file_direct_write(iocb, iter, pos);
 226 #endif /* CONFIG_NFS_SWAP */
 227 }
 228
 229 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 230 {
 231         unsigned int i;
 232         for (i = 0; i < npages; i++)
 233                 page_cache_release(pages[i]);
 234 }
 235
 236 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 237                               struct nfs_direct_req *dreq)
 238 {
 239         cinfo->lock = &dreq->lock;
 240         cinfo->mds = &dreq->mds_cinfo;
 241         cinfo->ds = &dreq->ds_cinfo;
 242         cinfo->dreq = dreq;
 243         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 244 }
 245
 246 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 247 {
 248         struct nfs_direct_req *dreq;
 249
 250         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 251         if (!dreq)
 252                 return NULL;
 253
 254         kref_init(&dreq->kref);
 255         kref_get(&dreq->kref);
 256         init_completion(&dreq->completion);
 257         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 258         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 259         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 260         spin_lock_init(&dreq->lock);
 261
 262         return dreq;
 263 }
 264
 265 static void nfs_direct_req_free(struct kref *kref)
 266 {
 267         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 268
 269         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 270         if (dreq->l_ctx != NULL)
 271                 nfs_put_lock_context(dreq->l_ctx);
 272         if (dreq->ctx != NULL)
 273                 put_nfs_open_context(dreq->ctx);
 274         kmem_cache_free(nfs_direct_cachep, dreq);
 275 }
 276
 277 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 278 {
 279         kref_put(&dreq->kref, nfs_direct_req_free);
 280 }
 281
 282 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 283 {
 284         return dreq->bytes_left;
 285 }
 286 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 287
 288 /*
 289  * Collects and returns the final error value/byte-count.
 290  */
 291 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 292 {
 293         ssize_t result = -EIOCBQUEUED;
 294
 295         /* Async requests don't wait here */
 296         if (dreq->iocb)
 297                 goto out;
 298
 299         result = wait_for_completion_killable(&dreq->completion);
 300
 301         if (!result)
 302                 result = dreq->error;
 303         if (!result)
 304                 result = dreq->count;
 305
 306 out:
 307         return (ssize_t) result;
 308 }
 309
 310 /*
 311  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 312  * the iocb is still valid here if this is a synchronous request.
 313  */
 314 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 315 {
 316         struct inode *inode = dreq->inode;
 317
 318         if (dreq->iocb && write) {
 319                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 320
 321                 spin_lock(&inode->i_lock);
 322                 if (i_size_read(inode) < pos)
 323                         i_size_write(inode, pos);
 324                 spin_unlock(&inode->i_lock);
 325         }
 326
 327         if (write)
 328                 nfs_zap_mapping(inode, inode->i_mapping);
 329
 330         inode_dio_done(inode);
 331
 332         if (dreq->iocb) {
 333                 long res = (long) dreq->error;
 334                 if (!res)
 335                         res = (long) dreq->count;
 336                 aio_complete(dreq->iocb, res, 0);
 337         }
 338
 339         complete_all(&dreq->completion);
 340
 341         nfs_direct_req_release(dreq);
 342 }
 343
 344 static void nfs_direct_readpage_release(struct nfs_page *req)
 345 {
 346         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 347                 req->wb_context->dentry->d_inode->i_sb->s_id,
 348                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 349                 req->wb_bytes,
 350                 (long long)req_offset(req));
 351         nfs_release_request(req);
 352 }
 353
 354 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 355 {
 356         unsigned long bytes = 0;
 357         struct nfs_direct_req *dreq = hdr->dreq;
 358
 359         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 360                 goto out_put;
 361
 362         spin_lock(&dreq->lock);
 363         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 364                 dreq->error = hdr->error;
 365         else
 366                 dreq->count += hdr->good_bytes;
 367         spin_unlock(&dreq->lock);
 368
 369         while (!list_empty(&hdr->pages)) {
 370                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 371                 struct page *page = req->wb_page;
 372
 373                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 374                         set_page_dirty(page);
 375                 bytes += req->wb_bytes;
 376                 nfs_list_remove_request(req);
 377                 nfs_direct_readpage_release(req);
 378         }
 379 out_put:
 380         if (put_dreq(dreq))
 381                 nfs_direct_complete(dreq, false);
 382         hdr->release(hdr);
 383 }
 384
 385 static void nfs_read_sync_pgio_error(struct list_head *head)
 386 {
 387         struct nfs_page *req;
 388
 389         while (!list_empty(head)) {
 390                 req = nfs_list_entry(head->next);
 391                 nfs_list_remove_request(req);
 392                 nfs_release_request(req);
 393         }
 394 }
 395
 396 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 397 {
 398         get_dreq(hdr->dreq);
 399 }
 400
 401 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 402         .error_cleanup = nfs_read_sync_pgio_error,
 403         .init_hdr = nfs_direct_pgio_init,
 404         .completion = nfs_direct_read_completion,
 405 };
 406
 407 /*
 408  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 409  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 410  * bail and stop sending more reads.  Read length accounting is
 411  * handled automatically by nfs_direct_read_result().  Otherwise, if
 412  * no requests have been sent, just return an error.
 413  */
 414
 415 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 416                                               struct iov_iter *iter,
 417                                               loff_t pos)
 418 {
 419         struct nfs_pageio_descriptor desc;
 420         struct inode *inode = dreq->inode;
 421         ssize_t result = -EINVAL;
 422         size_t requested_bytes = 0;
 423         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 424
 425         nfs_pageio_init_read(&desc, dreq->inode, false,
 426                              &nfs_direct_read_completion_ops);
 427         get_dreq(dreq);
 428         desc.pg_dreq = dreq;
 429         atomic_inc(&inode->i_dio_count);
 430
 431         while (iov_iter_count(iter)) {
 432                 struct page **pagevec;
 433                 size_t bytes;
 434                 size_t pgbase;
 435                 unsigned npages, i;
 436
 437                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 438                                                   rsize, &pgbase);
 439                 if (result < 0)
 440                         break;
 441
 442                 bytes = result;
 443                 iov_iter_advance(iter, bytes);
 444                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 445                 for (i = 0; i < npages; i++) {
 446                         struct nfs_page *req;
 447                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 448                         /* XXX do we need to do the eof zeroing found in async_filler? */
 449                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 450                                                  pgbase, req_len);
 451                         if (IS_ERR(req)) {
 452                                 result = PTR_ERR(req);
 453                                 break;
 454                         }
 455                         req->wb_index = pos >> PAGE_SHIFT;
 456                         req->wb_offset = pos & ~PAGE_MASK;
 457                         if (!nfs_pageio_add_request(&desc, req)) {
 458                                 result = desc.pg_error;
 459                                 nfs_release_request(req);
 460                                 break;
 461                         }
 462                         pgbase = 0;
 463                         bytes -= req_len;
 464                         requested_bytes += req_len;
 465                         pos += req_len;
 466                         dreq->bytes_left -= req_len;
 467                 }
 468                 nfs_direct_release_pages(pagevec, npages);
 469                 kvfree(pagevec);
 470                 if (result < 0)
 471                         break;
 472         }
 473
 474         nfs_pageio_complete(&desc);
 475
 476         /*
 477          * If no bytes were started, return the error, and let the
 478          * generic layer handle the completion.
 479          */
 480         if (requested_bytes == 0) {
 481                 inode_dio_done(inode);
 482                 nfs_direct_req_release(dreq);
 483                 return result < 0 ? result : -EIO;
 484         }
 485
 486         if (put_dreq(dreq))
 487                 nfs_direct_complete(dreq, false);
 488         return 0;
 489 }
 490
 491 /**
 492  * nfs_file_direct_read - file direct read operation for NFS files
 493  * @iocb: target I/O control block
 494  * @iter: vector of user buffers into which to read data
 495  * @pos: byte offset in file where reading starts
 496  *
 497  * We use this function for direct reads instead of calling
 498  * generic_file_aio_read() in order to avoid gfar's check to see if
 499  * the request starts before the end of the file.  For that check
 500  * to work, we must generate a GETATTR before each direct read, and
 501  * even then there is a window between the GETATTR and the subsequent
 502  * READ where the file size could change.  Our preference is simply
 503  * to do all reads the application wants, and the server will take
 504  * care of managing the end of file boundary.
 505  *
 506  * This function also eliminates unnecessarily updating the file's
 507  * atime locally, as the NFS server sets the file's atime, and this
 508  * client must read the updated atime from the server back into its
 509  * cache.
 510  */
 511 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 512                                 loff_t pos)
 513 {
 514         struct file *file = iocb->ki_filp;
 515         struct address_space *mapping = file->f_mapping;
 516         struct inode *inode = mapping->host;
 517         struct nfs_direct_req *dreq;
 518         struct nfs_lock_context *l_ctx;
 519         ssize_t result = -EINVAL;
 520         size_t count = iov_iter_count(iter);
 521         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 522
 523         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 524                 file, count, (long long) pos);
 525
 526         result = 0;
 527         if (!count)
 528                 goto out;
 529
 530         mutex_lock(&inode->i_mutex);
 531         result = nfs_sync_mapping(mapping);
 532         if (result)
 533                 goto out_unlock;
 534
 535         task_io_account_read(count);
 536
 537         result = -ENOMEM;
 538         dreq = nfs_direct_req_alloc();
 539         if (dreq == NULL)
 540                 goto out_unlock;
 541
 542         dreq->inode = inode;
 543         dreq->bytes_left = count;
 544         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 545         l_ctx = nfs_get_lock_context(dreq->ctx);
 546         if (IS_ERR(l_ctx)) {
 547                 result = PTR_ERR(l_ctx);
 548                 goto out_release;
 549         }
 550         dreq->l_ctx = l_ctx;
 551         if (!is_sync_kiocb(iocb))
 552                 dreq->iocb = iocb;
 553
 554         NFS_I(inode)->read_io += count;
 555         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 556
 557         mutex_unlock(&inode->i_mutex);
 558
 559         if (!result) {
 560                 result = nfs_direct_wait(dreq);
 561                 if (result > 0)
 562                         iocb->ki_pos = pos + result;
 563         }
 564
 565         nfs_direct_req_release(dreq);
 566         return result;
 567
 568 out_release:
 569         nfs_direct_req_release(dreq);
 570 out_unlock:
 571         mutex_unlock(&inode->i_mutex);
 572 out:
 573         return result;
 574 }
 575
 576 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 577 {
 578         struct nfs_pageio_descriptor desc;
 579         struct nfs_page *req, *tmp;
 580         LIST_HEAD(reqs);
 581         struct nfs_commit_info cinfo;
 582         LIST_HEAD(failed);
 583
 584         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 585         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 586         spin_lock(cinfo.lock);
 587         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 588         spin_unlock(cinfo.lock);
 589
 590         dreq->count = 0;
 591         get_dreq(dreq);
 592
 593         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 594                               &nfs_direct_write_completion_ops);
 595         desc.pg_dreq = dreq;
 596
 597         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 598                 if (!nfs_pageio_add_request(&desc, req)) {
 599                         nfs_list_remove_request(req);
 600                         nfs_list_add_request(req, &failed);
 601                         spin_lock(cinfo.lock);
 602                         dreq->flags = 0;
 603                         dreq->error = -EIO;
 604                         spin_unlock(cinfo.lock);
 605                 }
 606                 nfs_release_request(req);
 607         }
 608         nfs_pageio_complete(&desc);
 609
 610         while (!list_empty(&failed)) {
 611                 req = nfs_list_entry(failed.next);
 612                 nfs_list_remove_request(req);
 613                 nfs_unlock_and_release_request(req);
 614         }
 615
 616         if (put_dreq(dreq))
 617                 nfs_direct_write_complete(dreq, dreq->inode);
 618 }
 619
 620 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 621 {
 622         struct nfs_direct_req *dreq = data->dreq;
 623         struct nfs_commit_info cinfo;
 624         struct nfs_page *req;
 625         int status = data->task.tk_status;
 626
 627         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 628         if (status < 0) {
 629                 dprintk("NFS: %5u commit failed with error %d.\n",
 630                         data->task.tk_pid, status);
 631                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 632         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 633                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 634                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 635         }
 636
 637         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 638         while (!list_empty(&data->pages)) {
 639                 req = nfs_list_entry(data->pages.next);
 640                 nfs_list_remove_request(req);
 641                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 642                         /* Note the rewrite will go through mds */
 643                         nfs_mark_request_commit(req, NULL, &cinfo);
 644                 } else
 645                         nfs_release_request(req);
 646                 nfs_unlock_and_release_request(req);
 647         }
 648
 649         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 650                 nfs_direct_write_complete(dreq, data->inode);
 651 }
 652
 653 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 654 {
 655         /* There is no lock to clear */
 656 }
 657
 658 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 659         .completion = nfs_direct_commit_complete,
 660         .error_cleanup = nfs_direct_error_cleanup,
 661 };
 662
 663 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 664 {
 665         int res;
 666         struct nfs_commit_info cinfo;
 667         LIST_HEAD(mds_list);
 668
 669         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 670         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 671         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 672         if (res < 0) /* res == -ENOMEM */
 673                 nfs_direct_write_reschedule(dreq);
 674 }
 675
 676 static void nfs_direct_write_schedule_work(struct work_struct *work)
 677 {
 678         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 679         int flags = dreq->flags;
 680
 681         dreq->flags = 0;
 682         switch (flags) {
 683                 case NFS_ODIRECT_DO_COMMIT:
 684                         nfs_direct_commit_schedule(dreq);
 685                         break;
 686                 case NFS_ODIRECT_RESCHED_WRITES:
 687                         nfs_direct_write_reschedule(dreq);
 688                         break;
 689                 default:
 690                         nfs_direct_complete(dreq, true);
 691         }
 692 }
 693
 694 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 695 {
 696         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 697 }
 698
 699 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 700 {
 701         struct nfs_direct_req *dreq = hdr->dreq;
 702         struct nfs_commit_info cinfo;
 703         bool request_commit = false;
 704         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 705
 706         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 707                 goto out_put;
 708
 709         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 710
 711         spin_lock(&dreq->lock);
 712
 713         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 714                 dreq->flags = 0;
 715                 dreq->error = hdr->error;
 716         }
 717         if (dreq->error == 0) {
 718                 dreq->count += hdr->good_bytes;
 719                 if (nfs_write_need_commit(hdr)) {
 720                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 721                                 request_commit = true;
 722                         else if (dreq->flags == 0) {
 723                                 nfs_direct_set_hdr_verf(dreq, hdr);
 724                                 request_commit = true;
 725                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 726                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 727                                 request_commit = true;
 728                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 729                                         dreq->flags =
 730                                                 NFS_ODIRECT_RESCHED_WRITES;
 731                         }
 732                 }
 733         }
 734         spin_unlock(&dreq->lock);
 735
 736         while (!list_empty(&hdr->pages)) {
 737
 738                 req = nfs_list_entry(hdr->pages.next);
 739                 nfs_list_remove_request(req);
 740                 if (request_commit) {
 741                         kref_get(&req->wb_kref);
 742                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 743                 }
 744                 nfs_unlock_and_release_request(req);
 745         }
 746
 747 out_put:
 748         if (put_dreq(dreq))
 749                 nfs_direct_write_complete(dreq, hdr->inode);
 750         hdr->release(hdr);
 751 }
 752
 753 static void nfs_write_sync_pgio_error(struct list_head *head)
 754 {
 755         struct nfs_page *req;
 756
 757         while (!list_empty(head)) {
 758                 req = nfs_list_entry(head->next);
 759                 nfs_list_remove_request(req);
 760                 nfs_unlock_and_release_request(req);
 761         }
 762 }
 763
 764 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 765         .error_cleanup = nfs_write_sync_pgio_error,
 766         .init_hdr = nfs_direct_pgio_init,
 767         .completion = nfs_direct_write_completion,
 768 };
 769
 770
 771 /*
 772  * NB: Return the value of the first error return code.  Subsequent
 773  *     errors after the first one are ignored.
 774  */
 775 /*
 776  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 777  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 778  * bail and stop sending more writes.  Write length accounting is
 779  * handled automatically by nfs_direct_write_result().  Otherwise, if
 780  * no requests have been sent, just return an error.
 781  */
 782 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 783                                                struct iov_iter *iter,
 784                                                loff_t pos)
 785 {
 786         struct nfs_pageio_descriptor desc;
 787         struct inode *inode = dreq->inode;
 788         ssize_t result = 0;
 789         size_t requested_bytes = 0;
 790         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 791
 792         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 793                               &nfs_direct_write_completion_ops);
 794         desc.pg_dreq = dreq;
 795         get_dreq(dreq);
 796         atomic_inc(&inode->i_dio_count);
 797
 798         NFS_I(inode)->write_io += iov_iter_count(iter);
 799         while (iov_iter_count(iter)) {
 800                 struct page **pagevec;
 801                 size_t bytes;
 802                 size_t pgbase;
 803                 unsigned npages, i;
 804
 805                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 806                                                   wsize, &pgbase);
 807                 if (result < 0)
 808                         break;
 809
 810                 bytes = result;
 811                 iov_iter_advance(iter, bytes);
 812                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 813                 for (i = 0; i < npages; i++) {
 814                         struct nfs_page *req;
 815                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 816
 817                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 818                                                  pgbase, req_len);
 819                         if (IS_ERR(req)) {
 820                                 result = PTR_ERR(req);
 821                                 break;
 822                         }
 823                         nfs_lock_request(req);
 824                         req->wb_index = pos >> PAGE_SHIFT;
 825                         req->wb_offset = pos & ~PAGE_MASK;
 826                         if (!nfs_pageio_add_request(&desc, req)) {
 827                                 result = desc.pg_error;
 828                                 nfs_unlock_and_release_request(req);
 829                                 break;
 830                         }
 831                         pgbase = 0;
 832                         bytes -= req_len;
 833                         requested_bytes += req_len;
 834                         pos += req_len;
 835                         dreq->bytes_left -= req_len;
 836                 }
 837                 nfs_direct_release_pages(pagevec, npages);
 838                 kvfree(pagevec);
 839                 if (result < 0)
 840                         break;
 841         }
 842         nfs_pageio_complete(&desc);
 843
 844         /*
 845          * If no bytes were started, return the error, and let the
 846          * generic layer handle the completion.
 847          */
 848         if (requested_bytes == 0) {
 849                 inode_dio_done(inode);
 850                 nfs_direct_req_release(dreq);
 851                 return result < 0 ? result : -EIO;
 852         }
 853
 854         if (put_dreq(dreq))
 855                 nfs_direct_write_complete(dreq, dreq->inode);
 856         return 0;
 857 }
 858
 859 /**
 860  * nfs_file_direct_write - file direct write operation for NFS files
 861  * @iocb: target I/O control block
 862  * @iter: vector of user buffers from which to write data
 863  * @pos: byte offset in file where writing starts
 864  *
 865  * We use this function for direct writes instead of calling
 866  * generic_file_aio_write() in order to avoid taking the inode
 867  * semaphore and updating the i_size.  The NFS server will set
 868  * the new i_size and this client must read the updated size
 869  * back into its cache.  We let the server do generic write
 870  * parameter checking and report problems.
 871  *
 872  * We eliminate local atime updates, see direct read above.
 873  *
 874  * We avoid unnecessary page cache invalidations for normal cached
 875  * readers of this file.
 876  *
 877  * Note that O_APPEND is not supported for NFS direct writes, as there
 878  * is no atomic O_APPEND write facility in the NFS protocol.
 879  */
 880 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 881                                 loff_t pos)
 882 {
 883         ssize_t result = -EINVAL;
 884         struct file *file = iocb->ki_filp;
 885         struct address_space *mapping = file->f_mapping;
 886         struct inode *inode = mapping->host;
 887         struct nfs_direct_req *dreq;
 888         struct nfs_lock_context *l_ctx;
 889         loff_t end;
 890         size_t count = iov_iter_count(iter);
 891         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 892
 893         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 894
 895         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 896                 file, count, (long long) pos);
 897
 898         result = generic_write_checks(file, &pos, &count, 0);
 899         if (result)
 900                 goto out;
 901
 902         result = -EINVAL;
 903         if ((ssize_t) count < 0)
 904                 goto out;
 905         result = 0;
 906         if (!count)
 907                 goto out;
 908
 909         mutex_lock(&inode->i_mutex);
 910
 911         result = nfs_sync_mapping(mapping);
 912         if (result)
 913                 goto out_unlock;
 914
 915         if (mapping->nrpages) {
 916                 result = invalidate_inode_pages2_range(mapping,
 917                                         pos >> PAGE_CACHE_SHIFT, end);
 918                 if (result)
 919                         goto out_unlock;
 920         }
 921
 922         task_io_account_write(count);
 923
 924         result = -ENOMEM;
 925         dreq = nfs_direct_req_alloc();
 926         if (!dreq)
 927                 goto out_unlock;
 928
 929         dreq->inode = inode;
 930         dreq->bytes_left = count;
 931         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 932         l_ctx = nfs_get_lock_context(dreq->ctx);
 933         if (IS_ERR(l_ctx)) {
 934                 result = PTR_ERR(l_ctx);
 935                 goto out_release;
 936         }
 937         dreq->l_ctx = l_ctx;
 938         if (!is_sync_kiocb(iocb))
 939                 dreq->iocb = iocb;
 940
 941         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
 942
 943         if (mapping->nrpages) {
 944                 invalidate_inode_pages2_range(mapping,
 945                                               pos >> PAGE_CACHE_SHIFT, end);
 946         }
 947
 948         mutex_unlock(&inode->i_mutex);
 949
 950         if (!result) {
 951                 result = nfs_direct_wait(dreq);
 952                 if (result > 0) {
 953                         struct inode *inode = mapping->host;
 954
 955                         iocb->ki_pos = pos + result;
 956                         spin_lock(&inode->i_lock);
 957                         if (i_size_read(inode) < iocb->ki_pos)
 958                                 i_size_write(inode, iocb->ki_pos);
 959                         spin_unlock(&inode->i_lock);
 960                 }
 961         }
 962         nfs_direct_req_release(dreq);
 963         return result;
 964
 965 out_release:
 966         nfs_direct_req_release(dreq);
 967 out_unlock:
 968         mutex_unlock(&inode->i_mutex);
 969 out:
 970         return result;
 971 }
 972
 973 /**
 974  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 975  *
 976  */
 977 int __init nfs_init_directcache(void)
 978 {
 979         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 980                                                 sizeof(struct nfs_direct_req),
 981                                                 0, (SLAB_RECLAIM_ACCOUNT|
 982                                                         SLAB_MEM_SPREAD),
 983                                                 NULL);
 984         if (nfs_direct_cachep == NULL)
 985                 return -ENOMEM;
 986
 987         return 0;
 988 }
 989
 990 /**
 991  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 992  *
 993  */
 994 void nfs_destroy_directcache(void)
 995 {
 996         kmem_cache_destroy(nfs_direct_cachep);
 997 }