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_mirror {
  70         ssize_t count;
  71 };
  72
  73 struct nfs_direct_req {
  74         struct kref             kref;           /* release manager */
  75
  76         /* I/O parameters */
  77         struct nfs_open_context *ctx;           /* file open context info */
  78         struct nfs_lock_context *l_ctx;         /* Lock context info */
  79         struct kiocb *          iocb;           /* controlling i/o request */
  80         struct inode *          inode;          /* target file of i/o */
  81
  82         /* completion state */
  83         atomic_t                io_count;       /* i/os we're waiting for */
  84         spinlock_t              lock;           /* protect completion state */
  85
  86         struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
  87         int                     mirror_count;
  88
  89         ssize_t                 count,          /* bytes actually processed */
  90                                 bytes_left,     /* bytes left to be sent */
  91                                 io_start,       /* start of IO */
  92                                 error;          /* any reported error */
  93         struct completion       completion;     /* wait for i/o completion */
  94
  95         /* commit state */
  96         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  97         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  98         struct work_struct      work;
  99         int                     flags;
 100 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
 101 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
 102         struct nfs_writeverf    verf;           /* unstable write verifier */
 103 };
 104
 105 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 106 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 107 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
 108 static void nfs_direct_write_schedule_work(struct work_struct *work);
 109
 110 static inline void get_dreq(struct nfs_direct_req *dreq)
 111 {
 112         atomic_inc(&dreq->io_count);
 113 }
 114
 115 static inline int put_dreq(struct nfs_direct_req *dreq)
 116 {
 117         return atomic_dec_and_test(&dreq->io_count);
 118 }
 119
 120 void nfs_direct_set_resched_writes(struct nfs_direct_req *dreq)
 121 {
 122         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 123 }
 124 EXPORT_SYMBOL_GPL(nfs_direct_set_resched_writes);
 125
 126 static void
 127 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
 128 {
 129         int i;
 130         ssize_t count;
 131
 132         if (dreq->mirror_count == 1) {
 133                 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
 134                 dreq->count += hdr->good_bytes;
 135         } else {
 136                 /* mirrored writes */
 137                 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
 138                 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
 139                         count = hdr->io_start + hdr->good_bytes - dreq->io_start;
 140                         dreq->mirrors[hdr->pgio_mirror_idx].count = count;
 141                 }
 142                 /* update the dreq->count by finding the minimum agreed count from all
 143                  * mirrors */
 144                 count = dreq->mirrors[0].count;
 145
 146                 for (i = 1; i < dreq->mirror_count; i++)
 147                         count = min(count, dreq->mirrors[i].count);
 148
 149                 dreq->count = count;
 150         }
 151 }
 152
 153 /*
 154  * nfs_direct_select_verf - select the right verifier
 155  * @dreq - direct request possibly spanning multiple servers
 156  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 157  * @commit_idx - commit bucket index for the DS
 158  *
 159  * returns the correct verifier to use given the role of the server
 160  */
 161 static struct nfs_writeverf *
 162 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 163                        struct nfs_client *ds_clp,
 164                        int commit_idx)
 165 {
 166         struct nfs_writeverf *verfp = &dreq->verf;
 167
 168 #ifdef CONFIG_NFS_V4_1
 169         if (ds_clp) {
 170                 /* pNFS is in use, use the DS verf */
 171                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 172                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 173                 else
 174                         WARN_ON_ONCE(1);
 175         }
 176 #endif
 177         return verfp;
 178 }
 179
 180
 181 /*
 182  * nfs_direct_set_hdr_verf - set the write/commit verifier
 183  * @dreq - direct request possibly spanning multiple servers
 184  * @hdr - pageio header to validate against previously seen verfs
 185  *
 186  * Set the server's (MDS or DS) "seen" verifier
 187  */
 188 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 189                                     struct nfs_pgio_header *hdr)
 190 {
 191         struct nfs_writeverf *verfp;
 192
 193         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 194         WARN_ON_ONCE(verfp->committed >= 0);
 195         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 196         WARN_ON_ONCE(verfp->committed < 0);
 197 }
 198
 199 /*
 200  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 201  * @dreq - direct request possibly spanning multiple servers
 202  * @hdr - pageio header to validate against previously seen verf
 203  *
 204  * set the server's "seen" verf if not initialized.
 205  * returns result of comparison between @hdr->verf and the "seen"
 206  * verf of the server used by @hdr (DS or MDS)
 207  */
 208 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 209                                           struct nfs_pgio_header *hdr)
 210 {
 211         struct nfs_writeverf *verfp;
 212
 213         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 214         if (verfp->committed < 0) {
 215                 nfs_direct_set_hdr_verf(dreq, hdr);
 216                 return 0;
 217         }
 218         return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 219 }
 220
 221 /*
 222  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 223  * @dreq - direct request possibly spanning multiple servers
 224  * @data - commit data to validate against previously seen verf
 225  *
 226  * returns result of comparison between @data->verf and the verf of
 227  * the server used by @data (DS or MDS)
 228  */
 229 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 230                                            struct nfs_commit_data *data)
 231 {
 232         struct nfs_writeverf *verfp;
 233
 234         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 235                                          data->ds_commit_index);
 236
 237         /* verifier not set so always fail */
 238         if (verfp->committed < 0)
 239                 return 1;
 240
 241         return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
 242 }
 243
 244 /**
 245  * nfs_direct_IO - NFS address space operation for direct I/O
 246  * @rw: direction (read or write)
 247  * @iocb: target I/O control block
 248  * @iov: array of vectors that define I/O buffer
 249  * @pos: offset in file to begin the operation
 250  * @nr_segs: size of iovec array
 251  *
 252  * The presence of this routine in the address space ops vector means
 253  * the NFS client supports direct I/O. However, for most direct IO, we
 254  * shunt off direct read and write requests before the VFS gets them,
 255  * so this method is only ever called for swap.
 256  */
 257 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
 258 {
 259         struct inode *inode = iocb->ki_filp->f_mapping->host;
 260
 261         /* we only support swap file calling nfs_direct_IO */
 262         if (!IS_SWAPFILE(inode))
 263                 return 0;
 264
 265         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 266
 267         if (rw == READ)
 268                 return nfs_file_direct_read(iocb, iter, pos);
 269         return nfs_file_direct_write(iocb, iter, pos);
 270 }
 271
 272 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 273 {
 274         unsigned int i;
 275         for (i = 0; i < npages; i++)
 276                 page_cache_release(pages[i]);
 277 }
 278
 279 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 280                               struct nfs_direct_req *dreq)
 281 {
 282         cinfo->lock = &dreq->inode->i_lock;
 283         cinfo->mds = &dreq->mds_cinfo;
 284         cinfo->ds = &dreq->ds_cinfo;
 285         cinfo->dreq = dreq;
 286         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 287 }
 288
 289 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 290                                              struct nfs_pageio_descriptor *pgio,
 291                                              struct nfs_page *req)
 292 {
 293         int mirror_count = 1;
 294
 295         if (pgio->pg_ops->pg_get_mirror_count)
 296                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 297
 298         dreq->mirror_count = mirror_count;
 299 }
 300
 301 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 302 {
 303         struct nfs_direct_req *dreq;
 304
 305         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 306         if (!dreq)
 307                 return NULL;
 308
 309         kref_init(&dreq->kref);
 310         kref_get(&dreq->kref);
 311         init_completion(&dreq->completion);
 312         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 313         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 314         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 315         dreq->mirror_count = 1;
 316         spin_lock_init(&dreq->lock);
 317
 318         return dreq;
 319 }
 320
 321 static void nfs_direct_req_free(struct kref *kref)
 322 {
 323         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 324
 325         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 326         if (dreq->l_ctx != NULL)
 327                 nfs_put_lock_context(dreq->l_ctx);
 328         if (dreq->ctx != NULL)
 329                 put_nfs_open_context(dreq->ctx);
 330         kmem_cache_free(nfs_direct_cachep, dreq);
 331 }
 332
 333 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 334 {
 335         kref_put(&dreq->kref, nfs_direct_req_free);
 336 }
 337
 338 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 339 {
 340         return dreq->bytes_left;
 341 }
 342 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 343
 344 /*
 345  * Collects and returns the final error value/byte-count.
 346  */
 347 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 348 {
 349         ssize_t result = -EIOCBQUEUED;
 350
 351         /* Async requests don't wait here */
 352         if (dreq->iocb)
 353                 goto out;
 354
 355         result = wait_for_completion_killable(&dreq->completion);
 356
 357         if (!result)
 358                 result = dreq->error;
 359         if (!result)
 360                 result = dreq->count;
 361
 362 out:
 363         return (ssize_t) result;
 364 }
 365
 366 /*
 367  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 368  * the iocb is still valid here if this is a synchronous request.
 369  */
 370 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 371 {
 372         struct inode *inode = dreq->inode;
 373
 374         if (dreq->iocb && write) {
 375                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 376
 377                 spin_lock(&inode->i_lock);
 378                 if (i_size_read(inode) < pos)
 379                         i_size_write(inode, pos);
 380                 spin_unlock(&inode->i_lock);
 381         }
 382
 383         if (write)
 384                 nfs_zap_mapping(inode, inode->i_mapping);
 385
 386         inode_dio_done(inode);
 387
 388         if (dreq->iocb) {
 389                 long res = (long) dreq->error;
 390                 if (!res)
 391                         res = (long) dreq->count;
 392                 aio_complete(dreq->iocb, res, 0);
 393         }
 394
 395         complete_all(&dreq->completion);
 396
 397         nfs_direct_req_release(dreq);
 398 }
 399
 400 static void nfs_direct_readpage_release(struct nfs_page *req)
 401 {
 402         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 403                 req->wb_context->dentry->d_inode->i_sb->s_id,
 404                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 405                 req->wb_bytes,
 406                 (long long)req_offset(req));
 407         nfs_release_request(req);
 408 }
 409
 410 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 411 {
 412         unsigned long bytes = 0;
 413         struct nfs_direct_req *dreq = hdr->dreq;
 414
 415         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 416                 goto out_put;
 417
 418         spin_lock(&dreq->lock);
 419         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 420                 dreq->error = hdr->error;
 421         else
 422                 nfs_direct_good_bytes(dreq, hdr);
 423
 424         spin_unlock(&dreq->lock);
 425
 426         while (!list_empty(&hdr->pages)) {
 427                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 428                 struct page *page = req->wb_page;
 429
 430                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 431                         set_page_dirty(page);
 432                 bytes += req->wb_bytes;
 433                 nfs_list_remove_request(req);
 434                 nfs_direct_readpage_release(req);
 435         }
 436 out_put:
 437         if (put_dreq(dreq))
 438                 nfs_direct_complete(dreq, false);
 439         hdr->release(hdr);
 440 }
 441
 442 static void nfs_read_sync_pgio_error(struct list_head *head)
 443 {
 444         struct nfs_page *req;
 445
 446         while (!list_empty(head)) {
 447                 req = nfs_list_entry(head->next);
 448                 nfs_list_remove_request(req);
 449                 nfs_release_request(req);
 450         }
 451 }
 452
 453 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 454 {
 455         get_dreq(hdr->dreq);
 456 }
 457
 458 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 459         .error_cleanup = nfs_read_sync_pgio_error,
 460         .init_hdr = nfs_direct_pgio_init,
 461         .completion = nfs_direct_read_completion,
 462 };
 463
 464 /*
 465  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 466  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 467  * bail and stop sending more reads.  Read length accounting is
 468  * handled automatically by nfs_direct_read_result().  Otherwise, if
 469  * no requests have been sent, just return an error.
 470  */
 471
 472 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 473                                               struct iov_iter *iter,
 474                                               loff_t pos)
 475 {
 476         struct nfs_pageio_descriptor desc;
 477         struct inode *inode = dreq->inode;
 478         ssize_t result = -EINVAL;
 479         size_t requested_bytes = 0;
 480         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 481
 482         nfs_pageio_init_read(&desc, dreq->inode, false,
 483                              &nfs_direct_read_completion_ops);
 484         get_dreq(dreq);
 485         desc.pg_dreq = dreq;
 486         atomic_inc(&inode->i_dio_count);
 487
 488         while (iov_iter_count(iter)) {
 489                 struct page **pagevec;
 490                 size_t bytes;
 491                 size_t pgbase;
 492                 unsigned npages, i;
 493
 494                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 495                                                   rsize, &pgbase);
 496                 if (result < 0)
 497                         break;
 498
 499                 bytes = result;
 500                 iov_iter_advance(iter, bytes);
 501                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 502                 for (i = 0; i < npages; i++) {
 503                         struct nfs_page *req;
 504                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 505                         /* XXX do we need to do the eof zeroing found in async_filler? */
 506                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 507                                                  pgbase, req_len);
 508                         if (IS_ERR(req)) {
 509                                 result = PTR_ERR(req);
 510                                 break;
 511                         }
 512                         req->wb_index = pos >> PAGE_SHIFT;
 513                         req->wb_offset = pos & ~PAGE_MASK;
 514                         if (!nfs_pageio_add_request(&desc, req)) {
 515                                 result = desc.pg_error;
 516                                 nfs_release_request(req);
 517                                 break;
 518                         }
 519                         pgbase = 0;
 520                         bytes -= req_len;
 521                         requested_bytes += req_len;
 522                         pos += req_len;
 523                         dreq->bytes_left -= req_len;
 524                 }
 525                 nfs_direct_release_pages(pagevec, npages);
 526                 kvfree(pagevec);
 527                 if (result < 0)
 528                         break;
 529         }
 530
 531         nfs_pageio_complete(&desc);
 532
 533         /*
 534          * If no bytes were started, return the error, and let the
 535          * generic layer handle the completion.
 536          */
 537         if (requested_bytes == 0) {
 538                 inode_dio_done(inode);
 539                 nfs_direct_req_release(dreq);
 540                 return result < 0 ? result : -EIO;
 541         }
 542
 543         if (put_dreq(dreq))
 544                 nfs_direct_complete(dreq, false);
 545         return 0;
 546 }
 547
 548 /**
 549  * nfs_file_direct_read - file direct read operation for NFS files
 550  * @iocb: target I/O control block
 551  * @iter: vector of user buffers into which to read data
 552  * @pos: byte offset in file where reading starts
 553  *
 554  * We use this function for direct reads instead of calling
 555  * generic_file_aio_read() in order to avoid gfar's check to see if
 556  * the request starts before the end of the file.  For that check
 557  * to work, we must generate a GETATTR before each direct read, and
 558  * even then there is a window between the GETATTR and the subsequent
 559  * READ where the file size could change.  Our preference is simply
 560  * to do all reads the application wants, and the server will take
 561  * care of managing the end of file boundary.
 562  *
 563  * This function also eliminates unnecessarily updating the file's
 564  * atime locally, as the NFS server sets the file's atime, and this
 565  * client must read the updated atime from the server back into its
 566  * cache.
 567  */
 568 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 569                                 loff_t pos)
 570 {
 571         struct file *file = iocb->ki_filp;
 572         struct address_space *mapping = file->f_mapping;
 573         struct inode *inode = mapping->host;
 574         struct nfs_direct_req *dreq;
 575         struct nfs_lock_context *l_ctx;
 576         ssize_t result = -EINVAL;
 577         size_t count = iov_iter_count(iter);
 578         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 579
 580         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 581                 file, count, (long long) pos);
 582
 583         result = 0;
 584         if (!count)
 585                 goto out;
 586
 587         mutex_lock(&inode->i_mutex);
 588         result = nfs_sync_mapping(mapping);
 589         if (result)
 590                 goto out_unlock;
 591
 592         task_io_account_read(count);
 593
 594         result = -ENOMEM;
 595         dreq = nfs_direct_req_alloc();
 596         if (dreq == NULL)
 597                 goto out_unlock;
 598
 599         dreq->inode = inode;
 600         dreq->bytes_left = count;
 601         dreq->io_start = pos;
 602         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 603         l_ctx = nfs_get_lock_context(dreq->ctx);
 604         if (IS_ERR(l_ctx)) {
 605                 result = PTR_ERR(l_ctx);
 606                 goto out_release;
 607         }
 608         dreq->l_ctx = l_ctx;
 609         if (!is_sync_kiocb(iocb))
 610                 dreq->iocb = iocb;
 611
 612         NFS_I(inode)->read_io += count;
 613         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 614
 615         mutex_unlock(&inode->i_mutex);
 616
 617         if (!result) {
 618                 result = nfs_direct_wait(dreq);
 619                 if (result > 0)
 620                         iocb->ki_pos = pos + result;
 621         }
 622
 623         nfs_direct_req_release(dreq);
 624         return result;
 625
 626 out_release:
 627         nfs_direct_req_release(dreq);
 628 out_unlock:
 629         mutex_unlock(&inode->i_mutex);
 630 out:
 631         return result;
 632 }
 633
 634 static void
 635 nfs_direct_write_scan_commit_list(struct inode *inode,
 636                                   struct list_head *list,
 637                                   struct nfs_commit_info *cinfo)
 638 {
 639         spin_lock(cinfo->lock);
 640 #ifdef CONFIG_NFS_V4_1
 641         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 642                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 643 #endif
 644         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 645         spin_unlock(cinfo->lock);
 646 }
 647
 648 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 649 {
 650         struct nfs_pageio_descriptor desc;
 651         struct nfs_page *req, *tmp;
 652         LIST_HEAD(reqs);
 653         struct nfs_commit_info cinfo;
 654         LIST_HEAD(failed);
 655         int i;
 656
 657         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 658         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 659
 660         dreq->count = 0;
 661         for (i = 0; i < dreq->mirror_count; i++)
 662                 dreq->mirrors[i].count = 0;
 663         get_dreq(dreq);
 664
 665         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 666                               &nfs_direct_write_completion_ops);
 667         desc.pg_dreq = dreq;
 668
 669         req = nfs_list_entry(reqs.next);
 670         nfs_direct_setup_mirroring(dreq, &desc, req);
 671
 672         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 673                 if (!nfs_pageio_add_request(&desc, req)) {
 674                         nfs_list_remove_request(req);
 675                         nfs_list_add_request(req, &failed);
 676                         spin_lock(cinfo.lock);
 677                         dreq->flags = 0;
 678                         dreq->error = -EIO;
 679                         spin_unlock(cinfo.lock);
 680                 }
 681                 nfs_release_request(req);
 682         }
 683         nfs_pageio_complete(&desc);
 684
 685         while (!list_empty(&failed)) {
 686                 req = nfs_list_entry(failed.next);
 687                 nfs_list_remove_request(req);
 688                 nfs_unlock_and_release_request(req);
 689         }
 690
 691         if (put_dreq(dreq))
 692                 nfs_direct_write_complete(dreq, dreq->inode);
 693 }
 694
 695 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 696 {
 697         struct nfs_direct_req *dreq = data->dreq;
 698         struct nfs_commit_info cinfo;
 699         struct nfs_page *req;
 700         int status = data->task.tk_status;
 701
 702         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 703         if (status < 0) {
 704                 dprintk("NFS: %5u commit failed with error %d.\n",
 705                         data->task.tk_pid, status);
 706                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 707         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 708                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 709                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 710         }
 711
 712         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 713         while (!list_empty(&data->pages)) {
 714                 req = nfs_list_entry(data->pages.next);
 715                 nfs_list_remove_request(req);
 716                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 717                         /* Note the rewrite will go through mds */
 718                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 719                 } else
 720                         nfs_release_request(req);
 721                 nfs_unlock_and_release_request(req);
 722         }
 723
 724         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 725                 nfs_direct_write_complete(dreq, data->inode);
 726 }
 727
 728 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 729 {
 730         /* There is no lock to clear */
 731 }
 732
 733 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 734         .completion = nfs_direct_commit_complete,
 735         .error_cleanup = nfs_direct_error_cleanup,
 736 };
 737
 738 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 739 {
 740         int res;
 741         struct nfs_commit_info cinfo;
 742         LIST_HEAD(mds_list);
 743
 744         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 745         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 746         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 747         if (res < 0) /* res == -ENOMEM */
 748                 nfs_direct_write_reschedule(dreq);
 749 }
 750
 751 static void nfs_direct_write_schedule_work(struct work_struct *work)
 752 {
 753         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 754         int flags = dreq->flags;
 755
 756         dreq->flags = 0;
 757         switch (flags) {
 758                 case NFS_ODIRECT_DO_COMMIT:
 759                         nfs_direct_commit_schedule(dreq);
 760                         break;
 761                 case NFS_ODIRECT_RESCHED_WRITES:
 762                         nfs_direct_write_reschedule(dreq);
 763                         break;
 764                 default:
 765                         nfs_direct_complete(dreq, true);
 766         }
 767 }
 768
 769 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 770 {
 771         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 772 }
 773
 774 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 775 {
 776         struct nfs_direct_req *dreq = hdr->dreq;
 777         struct nfs_commit_info cinfo;
 778         bool request_commit = false;
 779         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 780
 781         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 782                 goto out_put;
 783
 784         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 785
 786         spin_lock(&dreq->lock);
 787
 788         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 789                 dreq->flags = 0;
 790                 dreq->error = hdr->error;
 791         }
 792         if (dreq->error == 0) {
 793                 nfs_direct_good_bytes(dreq, hdr);
 794                 if (nfs_write_need_commit(hdr)) {
 795                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 796                                 request_commit = true;
 797                         else if (dreq->flags == 0) {
 798                                 nfs_direct_set_hdr_verf(dreq, hdr);
 799                                 request_commit = true;
 800                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 801                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 802                                 request_commit = true;
 803                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 804                                         dreq->flags =
 805                                                 NFS_ODIRECT_RESCHED_WRITES;
 806                         }
 807                 }
 808         }
 809         spin_unlock(&dreq->lock);
 810
 811         while (!list_empty(&hdr->pages)) {
 812
 813                 req = nfs_list_entry(hdr->pages.next);
 814                 nfs_list_remove_request(req);
 815                 if (request_commit) {
 816                         kref_get(&req->wb_kref);
 817                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 818                                 hdr->ds_commit_idx);
 819                 }
 820                 nfs_unlock_and_release_request(req);
 821         }
 822
 823 out_put:
 824         if (put_dreq(dreq))
 825                 nfs_direct_write_complete(dreq, hdr->inode);
 826         hdr->release(hdr);
 827 }
 828
 829 static void nfs_write_sync_pgio_error(struct list_head *head)
 830 {
 831         struct nfs_page *req;
 832
 833         while (!list_empty(head)) {
 834                 req = nfs_list_entry(head->next);
 835                 nfs_list_remove_request(req);
 836                 nfs_unlock_and_release_request(req);
 837         }
 838 }
 839
 840 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 841         .error_cleanup = nfs_write_sync_pgio_error,
 842         .init_hdr = nfs_direct_pgio_init,
 843         .completion = nfs_direct_write_completion,
 844 };
 845
 846
 847 /*
 848  * NB: Return the value of the first error return code.  Subsequent
 849  *     errors after the first one are ignored.
 850  */
 851 /*
 852  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 853  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 854  * bail and stop sending more writes.  Write length accounting is
 855  * handled automatically by nfs_direct_write_result().  Otherwise, if
 856  * no requests have been sent, just return an error.
 857  */
 858 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 859                                                struct iov_iter *iter,
 860                                                loff_t pos)
 861 {
 862         struct nfs_pageio_descriptor desc;
 863         struct inode *inode = dreq->inode;
 864         ssize_t result = 0;
 865         size_t requested_bytes = 0;
 866         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 867
 868         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 869                               &nfs_direct_write_completion_ops);
 870         desc.pg_dreq = dreq;
 871         get_dreq(dreq);
 872         atomic_inc(&inode->i_dio_count);
 873
 874         NFS_I(inode)->write_io += iov_iter_count(iter);
 875         while (iov_iter_count(iter)) {
 876                 struct page **pagevec;
 877                 size_t bytes;
 878                 size_t pgbase;
 879                 unsigned npages, i;
 880
 881                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 882                                                   wsize, &pgbase);
 883                 if (result < 0)
 884                         break;
 885
 886                 bytes = result;
 887                 iov_iter_advance(iter, bytes);
 888                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 889                 for (i = 0; i < npages; i++) {
 890                         struct nfs_page *req;
 891                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 892
 893                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 894                                                  pgbase, req_len);
 895                         if (IS_ERR(req)) {
 896                                 result = PTR_ERR(req);
 897                                 break;
 898                         }
 899
 900                         nfs_direct_setup_mirroring(dreq, &desc, req);
 901
 902                         nfs_lock_request(req);
 903                         req->wb_index = pos >> PAGE_SHIFT;
 904                         req->wb_offset = pos & ~PAGE_MASK;
 905                         if (!nfs_pageio_add_request(&desc, req)) {
 906                                 result = desc.pg_error;
 907                                 nfs_unlock_and_release_request(req);
 908                                 break;
 909                         }
 910                         pgbase = 0;
 911                         bytes -= req_len;
 912                         requested_bytes += req_len;
 913                         pos += req_len;
 914                         dreq->bytes_left -= req_len;
 915                 }
 916                 nfs_direct_release_pages(pagevec, npages);
 917                 kvfree(pagevec);
 918                 if (result < 0)
 919                         break;
 920         }
 921         nfs_pageio_complete(&desc);
 922
 923         /*
 924          * If no bytes were started, return the error, and let the
 925          * generic layer handle the completion.
 926          */
 927         if (requested_bytes == 0) {
 928                 inode_dio_done(inode);
 929                 nfs_direct_req_release(dreq);
 930                 return result < 0 ? result : -EIO;
 931         }
 932
 933         if (put_dreq(dreq))
 934                 nfs_direct_write_complete(dreq, dreq->inode);
 935         return 0;
 936 }
 937
 938 /**
 939  * nfs_file_direct_write - file direct write operation for NFS files
 940  * @iocb: target I/O control block
 941  * @iter: vector of user buffers from which to write data
 942  * @pos: byte offset in file where writing starts
 943  *
 944  * We use this function for direct writes instead of calling
 945  * generic_file_aio_write() in order to avoid taking the inode
 946  * semaphore and updating the i_size.  The NFS server will set
 947  * the new i_size and this client must read the updated size
 948  * back into its cache.  We let the server do generic write
 949  * parameter checking and report problems.
 950  *
 951  * We eliminate local atime updates, see direct read above.
 952  *
 953  * We avoid unnecessary page cache invalidations for normal cached
 954  * readers of this file.
 955  *
 956  * Note that O_APPEND is not supported for NFS direct writes, as there
 957  * is no atomic O_APPEND write facility in the NFS protocol.
 958  */
 959 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 960                                 loff_t pos)
 961 {
 962         ssize_t result = -EINVAL;
 963         struct file *file = iocb->ki_filp;
 964         struct address_space *mapping = file->f_mapping;
 965         struct inode *inode = mapping->host;
 966         struct nfs_direct_req *dreq;
 967         struct nfs_lock_context *l_ctx;
 968         loff_t end;
 969         size_t count = iov_iter_count(iter);
 970         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 971
 972         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 973
 974         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 975                 file, count, (long long) pos);
 976
 977         result = generic_write_checks(file, &pos, &count, 0);
 978         if (result)
 979                 goto out;
 980
 981         result = -EINVAL;
 982         if ((ssize_t) count < 0)
 983                 goto out;
 984         result = 0;
 985         if (!count)
 986                 goto out;
 987
 988         mutex_lock(&inode->i_mutex);
 989
 990         result = nfs_sync_mapping(mapping);
 991         if (result)
 992                 goto out_unlock;
 993
 994         if (mapping->nrpages) {
 995                 result = invalidate_inode_pages2_range(mapping,
 996                                         pos >> PAGE_CACHE_SHIFT, end);
 997                 if (result)
 998                         goto out_unlock;
 999         }
1000
1001         task_io_account_write(count);
1002
1003         result = -ENOMEM;
1004         dreq = nfs_direct_req_alloc();
1005         if (!dreq)
1006                 goto out_unlock;
1007
1008         dreq->inode = inode;
1009         dreq->bytes_left = count;
1010         dreq->io_start = pos;
1011         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1012         l_ctx = nfs_get_lock_context(dreq->ctx);
1013         if (IS_ERR(l_ctx)) {
1014                 result = PTR_ERR(l_ctx);
1015                 goto out_release;
1016         }
1017         dreq->l_ctx = l_ctx;
1018         if (!is_sync_kiocb(iocb))
1019                 dreq->iocb = iocb;
1020
1021         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1022
1023         if (mapping->nrpages) {
1024                 invalidate_inode_pages2_range(mapping,
1025                                               pos >> PAGE_CACHE_SHIFT, end);
1026         }
1027
1028         mutex_unlock(&inode->i_mutex);
1029
1030         if (!result) {
1031                 result = nfs_direct_wait(dreq);
1032                 if (result > 0) {
1033                         struct inode *inode = mapping->host;
1034
1035                         iocb->ki_pos = pos + result;
1036                         spin_lock(&inode->i_lock);
1037                         if (i_size_read(inode) < iocb->ki_pos)
1038                                 i_size_write(inode, iocb->ki_pos);
1039                         spin_unlock(&inode->i_lock);
1040                         generic_write_sync(file, pos, result);
1041                 }
1042         }
1043         nfs_direct_req_release(dreq);
1044         return result;
1045
1046 out_release:
1047         nfs_direct_req_release(dreq);
1048 out_unlock:
1049         mutex_unlock(&inode->i_mutex);
1050 out:
1051         return result;
1052 }
1053
1054 /**
1055  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1056  *
1057  */
1058 int __init nfs_init_directcache(void)
1059 {
1060         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1061                                                 sizeof(struct nfs_direct_req),
1062                                                 0, (SLAB_RECLAIM_ACCOUNT|
1063                                                         SLAB_MEM_SPREAD),
1064                                                 NULL);
1065         if (nfs_direct_cachep == NULL)
1066                 return -ENOMEM;
1067
1068         return 0;
1069 }
1070
1071 /**
1072  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1073  *
1074  */
1075 void nfs_destroy_directcache(void)
1076 {
1077         kmem_cache_destroy(nfs_direct_cachep);
1078 }