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                                 max_count,      /* max expected count */
  91                                 bytes_left,     /* bytes left to be sent */
  92                                 io_start,       /* start of IO */
  93                                 error;          /* any reported error */
  94         struct completion       completion;     /* wait for i/o completion */
  95
  96         /* commit state */
  97         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  98         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  99         struct work_struct      work;
 100         int                     flags;
 101 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
 102 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
 103         struct nfs_writeverf    verf;           /* unstable write verifier */
 104 };
 105
 106 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 107 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 108 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
 109 static void nfs_direct_write_schedule_work(struct work_struct *work);
 110
 111 static inline void get_dreq(struct nfs_direct_req *dreq)
 112 {
 113         atomic_inc(&dreq->io_count);
 114 }
 115
 116 static inline int put_dreq(struct nfs_direct_req *dreq)
 117 {
 118         return atomic_dec_and_test(&dreq->io_count);
 119 }
 120
 121 static void
 122 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
 123 {
 124         int i;
 125         ssize_t count;
 126
 127         WARN_ON_ONCE(dreq->count >= dreq->max_count);
 128
 129         if (dreq->mirror_count == 1) {
 130                 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
 131                 dreq->count += hdr->good_bytes;
 132         } else {
 133                 /* mirrored writes */
 134                 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
 135                 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
 136                         count = hdr->io_start + hdr->good_bytes - dreq->io_start;
 137                         dreq->mirrors[hdr->pgio_mirror_idx].count = count;
 138                 }
 139                 /* update the dreq->count by finding the minimum agreed count from all
 140                  * mirrors */
 141                 count = dreq->mirrors[0].count;
 142
 143                 for (i = 1; i < dreq->mirror_count; i++)
 144                         count = min(count, dreq->mirrors[i].count);
 145
 146                 dreq->count = count;
 147         }
 148 }
 149
 150 /*
 151  * nfs_direct_select_verf - select the right verifier
 152  * @dreq - direct request possibly spanning multiple servers
 153  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 154  * @commit_idx - commit bucket index for the DS
 155  *
 156  * returns the correct verifier to use given the role of the server
 157  */
 158 static struct nfs_writeverf *
 159 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 160                        struct nfs_client *ds_clp,
 161                        int commit_idx)
 162 {
 163         struct nfs_writeverf *verfp = &dreq->verf;
 164
 165 #ifdef CONFIG_NFS_V4_1
 166         /*
 167          * pNFS is in use, use the DS verf except commit_through_mds is set
 168          * for layout segment where nbuckets is zero.
 169          */
 170         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 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  * @iocb: target I/O control block
 247  * @iter: I/O buffer
 248  *
 249  * The presence of this routine in the address space ops vector means
 250  * the NFS client supports direct I/O. However, for most direct IO, we
 251  * shunt off direct read and write requests before the VFS gets them,
 252  * so this method is only ever called for swap.
 253  */
 254 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 255 {
 256         struct inode *inode = iocb->ki_filp->f_mapping->host;
 257
 258         /* we only support swap file calling nfs_direct_IO */
 259         if (!IS_SWAPFILE(inode))
 260                 return 0;
 261
 262         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 263
 264         if (iov_iter_rw(iter) == READ)
 265                 return nfs_file_direct_read(iocb, iter);
 266         return nfs_file_direct_write(iocb, iter);
 267 }
 268
 269 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 270 {
 271         unsigned int i;
 272         for (i = 0; i < npages; i++)
 273                 put_page(pages[i]);
 274 }
 275
 276 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 277                               struct nfs_direct_req *dreq)
 278 {
 279         cinfo->inode = dreq->inode;
 280         cinfo->mds = &dreq->mds_cinfo;
 281         cinfo->ds = &dreq->ds_cinfo;
 282         cinfo->dreq = dreq;
 283         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 284 }
 285
 286 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 287                                              struct nfs_pageio_descriptor *pgio,
 288                                              struct nfs_page *req)
 289 {
 290         int mirror_count = 1;
 291
 292         if (pgio->pg_ops->pg_get_mirror_count)
 293                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 294
 295         dreq->mirror_count = mirror_count;
 296 }
 297
 298 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 299 {
 300         struct nfs_direct_req *dreq;
 301
 302         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 303         if (!dreq)
 304                 return NULL;
 305
 306         kref_init(&dreq->kref);
 307         kref_get(&dreq->kref);
 308         init_completion(&dreq->completion);
 309         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 310         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 311         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 312         dreq->mirror_count = 1;
 313         spin_lock_init(&dreq->lock);
 314
 315         return dreq;
 316 }
 317
 318 static void nfs_direct_req_free(struct kref *kref)
 319 {
 320         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 321
 322         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 323         if (dreq->l_ctx != NULL)
 324                 nfs_put_lock_context(dreq->l_ctx);
 325         if (dreq->ctx != NULL)
 326                 put_nfs_open_context(dreq->ctx);
 327         kmem_cache_free(nfs_direct_cachep, dreq);
 328 }
 329
 330 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 331 {
 332         kref_put(&dreq->kref, nfs_direct_req_free);
 333 }
 334
 335 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 336 {
 337         return dreq->bytes_left;
 338 }
 339 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 340
 341 /*
 342  * Collects and returns the final error value/byte-count.
 343  */
 344 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 345 {
 346         ssize_t result = -EIOCBQUEUED;
 347
 348         /* Async requests don't wait here */
 349         if (dreq->iocb)
 350                 goto out;
 351
 352         result = wait_for_completion_killable(&dreq->completion);
 353
 354         if (!result) {
 355                 result = dreq->count;
 356                 WARN_ON_ONCE(dreq->count < 0);
 357         }
 358         if (!result)
 359                 result = dreq->error;
 360
 361 out:
 362         return (ssize_t) result;
 363 }
 364
 365 /*
 366  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 367  * the iocb is still valid here if this is a synchronous request.
 368  */
 369 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 370 {
 371         struct inode *inode = dreq->inode;
 372
 373         if (dreq->iocb && write) {
 374                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 375
 376                 spin_lock(&inode->i_lock);
 377                 if (i_size_read(inode) < pos)
 378                         i_size_write(inode, pos);
 379                 spin_unlock(&inode->i_lock);
 380         }
 381
 382         if (write)
 383                 nfs_zap_mapping(inode, inode->i_mapping);
 384
 385         inode_dio_end(inode);
 386
 387         if (dreq->iocb) {
 388                 long res = (long) dreq->error;
 389                 if (dreq->count != 0) {
 390                         res = (long) dreq->count;
 391                         WARN_ON_ONCE(dreq->count < 0);
 392                 }
 393                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 394         }
 395
 396         complete_all(&dreq->completion);
 397
 398         nfs_direct_req_release(dreq);
 399 }
 400
 401 static void nfs_direct_readpage_release(struct nfs_page *req)
 402 {
 403         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 404                 req->wb_context->dentry->d_sb->s_id,
 405                 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
 406                 req->wb_bytes,
 407                 (long long)req_offset(req));
 408         nfs_release_request(req);
 409 }
 410
 411 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 412 {
 413         unsigned long bytes = 0;
 414         struct nfs_direct_req *dreq = hdr->dreq;
 415
 416         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 417                 goto out_put;
 418
 419         spin_lock(&dreq->lock);
 420         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 421                 dreq->error = hdr->error;
 422         else
 423                 nfs_direct_good_bytes(dreq, hdr);
 424
 425         spin_unlock(&dreq->lock);
 426
 427         while (!list_empty(&hdr->pages)) {
 428                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 429                 struct page *page = req->wb_page;
 430
 431                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 432                         set_page_dirty(page);
 433                 bytes += req->wb_bytes;
 434                 nfs_list_remove_request(req);
 435                 nfs_direct_readpage_release(req);
 436         }
 437 out_put:
 438         if (put_dreq(dreq))
 439                 nfs_direct_complete(dreq, false);
 440         hdr->release(hdr);
 441 }
 442
 443 static void nfs_read_sync_pgio_error(struct list_head *head)
 444 {
 445         struct nfs_page *req;
 446
 447         while (!list_empty(head)) {
 448                 req = nfs_list_entry(head->next);
 449                 nfs_list_remove_request(req);
 450                 nfs_release_request(req);
 451         }
 452 }
 453
 454 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 455 {
 456         get_dreq(hdr->dreq);
 457 }
 458
 459 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 460         .error_cleanup = nfs_read_sync_pgio_error,
 461         .init_hdr = nfs_direct_pgio_init,
 462         .completion = nfs_direct_read_completion,
 463 };
 464
 465 /*
 466  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 467  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 468  * bail and stop sending more reads.  Read length accounting is
 469  * handled automatically by nfs_direct_read_result().  Otherwise, if
 470  * no requests have been sent, just return an error.
 471  */
 472
 473 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 474                                               struct iov_iter *iter,
 475                                               loff_t pos)
 476 {
 477         struct nfs_pageio_descriptor desc;
 478         struct inode *inode = dreq->inode;
 479         ssize_t result = -EINVAL;
 480         size_t requested_bytes = 0;
 481         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 482
 483         nfs_pageio_init_read(&desc, dreq->inode, false,
 484                              &nfs_direct_read_completion_ops);
 485         get_dreq(dreq);
 486         desc.pg_dreq = dreq;
 487         inode_dio_begin(inode);
 488
 489         while (iov_iter_count(iter)) {
 490                 struct page **pagevec;
 491                 size_t bytes;
 492                 size_t pgbase;
 493                 unsigned npages, i;
 494
 495                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 496                                                   rsize, &pgbase);
 497                 if (result < 0)
 498                         break;
 499
 500                 bytes = result;
 501                 iov_iter_advance(iter, bytes);
 502                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 503                 for (i = 0; i < npages; i++) {
 504                         struct nfs_page *req;
 505                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 506                         /* XXX do we need to do the eof zeroing found in async_filler? */
 507                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 508                                                  pgbase, req_len);
 509                         if (IS_ERR(req)) {
 510                                 result = PTR_ERR(req);
 511                                 break;
 512                         }
 513                         req->wb_index = pos >> PAGE_SHIFT;
 514                         req->wb_offset = pos & ~PAGE_MASK;
 515                         if (!nfs_pageio_add_request(&desc, req)) {
 516                                 result = desc.pg_error;
 517                                 nfs_release_request(req);
 518                                 break;
 519                         }
 520                         pgbase = 0;
 521                         bytes -= req_len;
 522                         requested_bytes += req_len;
 523                         pos += req_len;
 524                         dreq->bytes_left -= req_len;
 525                 }
 526                 nfs_direct_release_pages(pagevec, npages);
 527                 kvfree(pagevec);
 528                 if (result < 0)
 529                         break;
 530         }
 531
 532         nfs_pageio_complete(&desc);
 533
 534         /*
 535          * If no bytes were started, return the error, and let the
 536          * generic layer handle the completion.
 537          */
 538         if (requested_bytes == 0) {
 539                 inode_dio_end(inode);
 540                 nfs_direct_req_release(dreq);
 541                 return result < 0 ? result : -EIO;
 542         }
 543
 544         if (put_dreq(dreq))
 545                 nfs_direct_complete(dreq, false);
 546         return 0;
 547 }
 548
 549 /**
 550  * nfs_file_direct_read - file direct read operation for NFS files
 551  * @iocb: target I/O control block
 552  * @iter: vector of user buffers into which to read data
 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 {
 570         struct file *file = iocb->ki_filp;
 571         struct address_space *mapping = file->f_mapping;
 572         struct inode *inode = mapping->host;
 573         struct nfs_direct_req *dreq;
 574         struct nfs_lock_context *l_ctx;
 575         ssize_t result = -EINVAL;
 576         size_t count = iov_iter_count(iter);
 577         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 578
 579         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 580                 file, count, (long long) iocb->ki_pos);
 581
 582         result = 0;
 583         if (!count)
 584                 goto out;
 585
 586         inode_lock(inode);
 587         result = nfs_sync_mapping(mapping);
 588         if (result)
 589                 goto out_unlock;
 590
 591         task_io_account_read(count);
 592
 593         result = -ENOMEM;
 594         dreq = nfs_direct_req_alloc();
 595         if (dreq == NULL)
 596                 goto out_unlock;
 597
 598         dreq->inode = inode;
 599         dreq->bytes_left = dreq->max_count = count;
 600         dreq->io_start = iocb->ki_pos;
 601         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 602         l_ctx = nfs_get_lock_context(dreq->ctx);
 603         if (IS_ERR(l_ctx)) {
 604                 result = PTR_ERR(l_ctx);
 605                 goto out_release;
 606         }
 607         dreq->l_ctx = l_ctx;
 608         if (!is_sync_kiocb(iocb))
 609                 dreq->iocb = iocb;
 610
 611         NFS_I(inode)->read_io += count;
 612         result = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 613
 614         inode_unlock(inode);
 615
 616         if (!result) {
 617                 result = nfs_direct_wait(dreq);
 618                 if (result > 0)
 619                         iocb->ki_pos += result;
 620         }
 621
 622         nfs_direct_req_release(dreq);
 623         return result;
 624
 625 out_release:
 626         nfs_direct_req_release(dreq);
 627 out_unlock:
 628         inode_unlock(inode);
 629 out:
 630         return result;
 631 }
 632
 633 static void
 634 nfs_direct_write_scan_commit_list(struct inode *inode,
 635                                   struct list_head *list,
 636                                   struct nfs_commit_info *cinfo)
 637 {
 638         spin_lock(&cinfo->inode->i_lock);
 639 #ifdef CONFIG_NFS_V4_1
 640         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 641                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 642 #endif
 643         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 644         spin_unlock(&cinfo->inode->i_lock);
 645 }
 646
 647 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 648 {
 649         struct nfs_pageio_descriptor desc;
 650         struct nfs_page *req, *tmp;
 651         LIST_HEAD(reqs);
 652         struct nfs_commit_info cinfo;
 653         LIST_HEAD(failed);
 654         int i;
 655
 656         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 657         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 658
 659         dreq->count = 0;
 660         for (i = 0; i < dreq->mirror_count; i++)
 661                 dreq->mirrors[i].count = 0;
 662         get_dreq(dreq);
 663
 664         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 665                               &nfs_direct_write_completion_ops);
 666         desc.pg_dreq = dreq;
 667
 668         req = nfs_list_entry(reqs.next);
 669         nfs_direct_setup_mirroring(dreq, &desc, req);
 670         if (desc.pg_error < 0) {
 671                 list_splice_init(&reqs, &failed);
 672                 goto out_failed;
 673         }
 674
 675         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 676                 if (!nfs_pageio_add_request(&desc, req)) {
 677                         nfs_list_remove_request(req);
 678                         nfs_list_add_request(req, &failed);
 679                         spin_lock(&cinfo.inode->i_lock);
 680                         dreq->flags = 0;
 681                         if (desc.pg_error < 0)
 682                                 dreq->error = desc.pg_error;
 683                         else
 684                                 dreq->error = -EIO;
 685                         spin_unlock(&cinfo.inode->i_lock);
 686                 }
 687                 nfs_release_request(req);
 688         }
 689         nfs_pageio_complete(&desc);
 690
 691 out_failed:
 692         while (!list_empty(&failed)) {
 693                 req = nfs_list_entry(failed.next);
 694                 nfs_list_remove_request(req);
 695                 nfs_unlock_and_release_request(req);
 696         }
 697
 698         if (put_dreq(dreq))
 699                 nfs_direct_write_complete(dreq, dreq->inode);
 700 }
 701
 702 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 703 {
 704         struct nfs_direct_req *dreq = data->dreq;
 705         struct nfs_commit_info cinfo;
 706         struct nfs_page *req;
 707         int status = data->task.tk_status;
 708
 709         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 710         if (status < 0) {
 711                 dprintk("NFS: %5u commit failed with error %d.\n",
 712                         data->task.tk_pid, status);
 713                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 714         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 715                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 716                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 717         }
 718
 719         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 720         while (!list_empty(&data->pages)) {
 721                 req = nfs_list_entry(data->pages.next);
 722                 nfs_list_remove_request(req);
 723                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 724                         /* Note the rewrite will go through mds */
 725                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 726                 } else
 727                         nfs_release_request(req);
 728                 nfs_unlock_and_release_request(req);
 729         }
 730
 731         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 732                 nfs_direct_write_complete(dreq, data->inode);
 733 }
 734
 735 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 736                 struct nfs_page *req)
 737 {
 738         struct nfs_direct_req *dreq = cinfo->dreq;
 739
 740         spin_lock(&dreq->lock);
 741         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 742         spin_unlock(&dreq->lock);
 743         nfs_mark_request_commit(req, NULL, cinfo, 0);
 744 }
 745
 746 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 747         .completion = nfs_direct_commit_complete,
 748         .resched_write = nfs_direct_resched_write,
 749 };
 750
 751 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 752 {
 753         int res;
 754         struct nfs_commit_info cinfo;
 755         LIST_HEAD(mds_list);
 756
 757         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 758         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 759         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 760         if (res < 0) /* res == -ENOMEM */
 761                 nfs_direct_write_reschedule(dreq);
 762 }
 763
 764 static void nfs_direct_write_schedule_work(struct work_struct *work)
 765 {
 766         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 767         int flags = dreq->flags;
 768
 769         dreq->flags = 0;
 770         switch (flags) {
 771                 case NFS_ODIRECT_DO_COMMIT:
 772                         nfs_direct_commit_schedule(dreq);
 773                         break;
 774                 case NFS_ODIRECT_RESCHED_WRITES:
 775                         nfs_direct_write_reschedule(dreq);
 776                         break;
 777                 default:
 778                         nfs_direct_complete(dreq, true);
 779         }
 780 }
 781
 782 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 783 {
 784         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 785 }
 786
 787 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 788 {
 789         struct nfs_direct_req *dreq = hdr->dreq;
 790         struct nfs_commit_info cinfo;
 791         bool request_commit = false;
 792         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 793
 794         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 795                 goto out_put;
 796
 797         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 798
 799         spin_lock(&dreq->lock);
 800
 801         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 802                 dreq->flags = 0;
 803                 dreq->error = hdr->error;
 804         }
 805         if (dreq->error == 0) {
 806                 nfs_direct_good_bytes(dreq, hdr);
 807                 if (nfs_write_need_commit(hdr)) {
 808                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 809                                 request_commit = true;
 810                         else if (dreq->flags == 0) {
 811                                 nfs_direct_set_hdr_verf(dreq, hdr);
 812                                 request_commit = true;
 813                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 814                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 815                                 request_commit = true;
 816                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 817                                         dreq->flags =
 818                                                 NFS_ODIRECT_RESCHED_WRITES;
 819                         }
 820                 }
 821         }
 822         spin_unlock(&dreq->lock);
 823
 824         while (!list_empty(&hdr->pages)) {
 825
 826                 req = nfs_list_entry(hdr->pages.next);
 827                 nfs_list_remove_request(req);
 828                 if (request_commit) {
 829                         kref_get(&req->wb_kref);
 830                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 831                                 hdr->ds_commit_idx);
 832                 }
 833                 nfs_unlock_and_release_request(req);
 834         }
 835
 836 out_put:
 837         if (put_dreq(dreq))
 838                 nfs_direct_write_complete(dreq, hdr->inode);
 839         hdr->release(hdr);
 840 }
 841
 842 static void nfs_write_sync_pgio_error(struct list_head *head)
 843 {
 844         struct nfs_page *req;
 845
 846         while (!list_empty(head)) {
 847                 req = nfs_list_entry(head->next);
 848                 nfs_list_remove_request(req);
 849                 nfs_unlock_and_release_request(req);
 850         }
 851 }
 852
 853 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 854 {
 855         struct nfs_direct_req *dreq = hdr->dreq;
 856
 857         spin_lock(&dreq->lock);
 858         if (dreq->error == 0) {
 859                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 860                 /* fake unstable write to let common nfs resend pages */
 861                 hdr->verf.committed = NFS_UNSTABLE;
 862                 hdr->good_bytes = hdr->args.count;
 863         }
 864         spin_unlock(&dreq->lock);
 865 }
 866
 867 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 868         .error_cleanup = nfs_write_sync_pgio_error,
 869         .init_hdr = nfs_direct_pgio_init,
 870         .completion = nfs_direct_write_completion,
 871         .reschedule_io = nfs_direct_write_reschedule_io,
 872 };
 873
 874
 875 /*
 876  * NB: Return the value of the first error return code.  Subsequent
 877  *     errors after the first one are ignored.
 878  */
 879 /*
 880  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 881  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 882  * bail and stop sending more writes.  Write length accounting is
 883  * handled automatically by nfs_direct_write_result().  Otherwise, if
 884  * no requests have been sent, just return an error.
 885  */
 886 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 887                                                struct iov_iter *iter,
 888                                                loff_t pos)
 889 {
 890         struct nfs_pageio_descriptor desc;
 891         struct inode *inode = dreq->inode;
 892         ssize_t result = 0;
 893         size_t requested_bytes = 0;
 894         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 895
 896         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 897                               &nfs_direct_write_completion_ops);
 898         desc.pg_dreq = dreq;
 899         get_dreq(dreq);
 900         inode_dio_begin(inode);
 901
 902         NFS_I(inode)->write_io += iov_iter_count(iter);
 903         while (iov_iter_count(iter)) {
 904                 struct page **pagevec;
 905                 size_t bytes;
 906                 size_t pgbase;
 907                 unsigned npages, i;
 908
 909                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 910                                                   wsize, &pgbase);
 911                 if (result < 0)
 912                         break;
 913
 914                 bytes = result;
 915                 iov_iter_advance(iter, bytes);
 916                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 917                 for (i = 0; i < npages; i++) {
 918                         struct nfs_page *req;
 919                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 920
 921                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 922                                                  pgbase, req_len);
 923                         if (IS_ERR(req)) {
 924                                 result = PTR_ERR(req);
 925                                 break;
 926                         }
 927
 928                         nfs_direct_setup_mirroring(dreq, &desc, req);
 929                         if (desc.pg_error < 0) {
 930                                 nfs_free_request(req);
 931                                 result = desc.pg_error;
 932                                 break;
 933                         }
 934
 935                         nfs_lock_request(req);
 936                         req->wb_index = pos >> PAGE_SHIFT;
 937                         req->wb_offset = pos & ~PAGE_MASK;
 938                         if (!nfs_pageio_add_request(&desc, req)) {
 939                                 result = desc.pg_error;
 940                                 nfs_unlock_and_release_request(req);
 941                                 break;
 942                         }
 943                         pgbase = 0;
 944                         bytes -= req_len;
 945                         requested_bytes += req_len;
 946                         pos += req_len;
 947                         dreq->bytes_left -= req_len;
 948                 }
 949                 nfs_direct_release_pages(pagevec, npages);
 950                 kvfree(pagevec);
 951                 if (result < 0)
 952                         break;
 953         }
 954         nfs_pageio_complete(&desc);
 955
 956         /*
 957          * If no bytes were started, return the error, and let the
 958          * generic layer handle the completion.
 959          */
 960         if (requested_bytes == 0) {
 961                 inode_dio_end(inode);
 962                 nfs_direct_req_release(dreq);
 963                 return result < 0 ? result : -EIO;
 964         }
 965
 966         if (put_dreq(dreq))
 967                 nfs_direct_write_complete(dreq, dreq->inode);
 968         return 0;
 969 }
 970
 971 /**
 972  * nfs_file_direct_write - file direct write operation for NFS files
 973  * @iocb: target I/O control block
 974  * @iter: vector of user buffers from which to write data
 975  *
 976  * We use this function for direct writes instead of calling
 977  * generic_file_aio_write() in order to avoid taking the inode
 978  * semaphore and updating the i_size.  The NFS server will set
 979  * the new i_size and this client must read the updated size
 980  * back into its cache.  We let the server do generic write
 981  * parameter checking and report problems.
 982  *
 983  * We eliminate local atime updates, see direct read above.
 984  *
 985  * We avoid unnecessary page cache invalidations for normal cached
 986  * readers of this file.
 987  *
 988  * Note that O_APPEND is not supported for NFS direct writes, as there
 989  * is no atomic O_APPEND write facility in the NFS protocol.
 990  */
 991 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 992 {
 993         ssize_t result = -EINVAL;
 994         struct file *file = iocb->ki_filp;
 995         struct address_space *mapping = file->f_mapping;
 996         struct inode *inode = mapping->host;
 997         struct nfs_direct_req *dreq;
 998         struct nfs_lock_context *l_ctx;
 999         loff_t pos, end;
1000
1001         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1002                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
1003
1004         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES,
1005                       iov_iter_count(iter));
1006
1007         pos = iocb->ki_pos;
1008         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1009
1010         inode_lock(inode);
1011
1012         result = nfs_sync_mapping(mapping);
1013         if (result)
1014                 goto out_unlock;
1015
1016         if (mapping->nrpages) {
1017                 result = invalidate_inode_pages2_range(mapping,
1018                                         pos >> PAGE_SHIFT, end);
1019                 if (result)
1020                         goto out_unlock;
1021         }
1022
1023         task_io_account_write(iov_iter_count(iter));
1024
1025         result = -ENOMEM;
1026         dreq = nfs_direct_req_alloc();
1027         if (!dreq)
1028                 goto out_unlock;
1029
1030         dreq->inode = inode;
1031         dreq->bytes_left = dreq->max_count = iov_iter_count(iter);
1032         dreq->io_start = pos;
1033         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1034         l_ctx = nfs_get_lock_context(dreq->ctx);
1035         if (IS_ERR(l_ctx)) {
1036                 result = PTR_ERR(l_ctx);
1037                 goto out_release;
1038         }
1039         dreq->l_ctx = l_ctx;
1040         if (!is_sync_kiocb(iocb))
1041                 dreq->iocb = iocb;
1042
1043         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1044
1045         if (mapping->nrpages) {
1046                 invalidate_inode_pages2_range(mapping,
1047                                               pos >> PAGE_SHIFT, end);
1048         }
1049
1050         inode_unlock(inode);
1051
1052         if (!result) {
1053                 result = nfs_direct_wait(dreq);
1054                 if (result > 0) {
1055                         struct inode *inode = mapping->host;
1056
1057                         iocb->ki_pos = pos + result;
1058                         spin_lock(&inode->i_lock);
1059                         if (i_size_read(inode) < iocb->ki_pos)
1060                                 i_size_write(inode, iocb->ki_pos);
1061                         spin_unlock(&inode->i_lock);
1062
1063                         /* XXX: should check the generic_write_sync retval */
1064                         generic_write_sync(iocb, result);
1065                 }
1066         }
1067         nfs_direct_req_release(dreq);
1068         return result;
1069
1070 out_release:
1071         nfs_direct_req_release(dreq);
1072 out_unlock:
1073         inode_unlock(inode);
1074         return result;
1075 }
1076
1077 /**
1078  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1079  *
1080  */
1081 int __init nfs_init_directcache(void)
1082 {
1083         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1084                                                 sizeof(struct nfs_direct_req),
1085                                                 0, (SLAB_RECLAIM_ACCOUNT|
1086                                                         SLAB_MEM_SPREAD),
1087                                                 NULL);
1088         if (nfs_direct_cachep == NULL)
1089                 return -ENOMEM;
1090
1091         return 0;
1092 }
1093
1094 /**
1095  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1096  *
1097  */
1098 void nfs_destroy_directcache(void)
1099 {
1100         kmem_cache_destroy(nfs_direct_cachep);
1101 }