fs/nfs/direct.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/direct.c
   4  *
   5  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   6  *
   7  * High-performance uncached I/O for the Linux NFS client
   8  *
   9  * There are important applications whose performance or correctness
  10  * depends on uncached access to file data.  Database clusters
  11  * (multiple copies of the same instance running on separate hosts)
  12  * implement their own cache coherency protocol that subsumes file
  13  * system cache protocols.  Applications that process datasets
  14  * considerably larger than the client's memory do not always benefit
  15  * from a local cache.  A streaming video server, for instance, has no
  16  * need to cache the contents of a file.
  17  *
  18  * When an application requests uncached I/O, all read and write requests
  19  * are made directly to the server; data stored or fetched via these
  20  * requests is not cached in the Linux page cache.  The client does not
  21  * correct unaligned requests from applications.  All requested bytes are
  22  * held on permanent storage before a direct write system call returns to
  23  * an application.
  24  *
  25  * Solaris implements an uncached I/O facility called directio() that
  26  * is used for backups and sequential I/O to very large files.  Solaris
  27  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  28  * an undocumented mount option.
  29  *
  30  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  31  * help from Andrew Morton.
  32  *
  33  * 18 Dec 2001  Initial implementation for 2.4  --cel
  34  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  35  * 08 Jun 2003  Port to 2.5 APIs  --cel
  36  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  37  * 15 Sep 2004  Parallel async reads  --cel
  38  * 04 May 2005  support O_DIRECT with aio  --cel
  39  *
  40  */
  41
  42 #include <linux/errno.h>
  43 #include <linux/sched.h>
  44 #include <linux/kernel.h>
  45 #include <linux/file.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/kref.h>
  48 #include <linux/slab.h>
  49 #include <linux/task_io_accounting_ops.h>
  50 #include <linux/module.h>
  51
  52 #include <linux/nfs_fs.h>
  53 #include <linux/nfs_page.h>
  54 #include <linux/sunrpc/clnt.h>
  55
  56 #include <linux/uaccess.h>
  57 #include <linux/atomic.h>
  58
  59 #include "internal.h"
  60 #include "iostat.h"
  61 #include "pnfs.h"
  62
  63 #define NFSDBG_FACILITY         NFSDBG_VFS
  64
  65 static struct kmem_cache *nfs_direct_cachep;
  66
  67 struct nfs_direct_req {
  68         struct kref             kref;           /* release manager */
  69
  70         /* I/O parameters */
  71         struct nfs_open_context *ctx;           /* file open context info */
  72         struct nfs_lock_context *l_ctx;         /* Lock context info */
  73         struct kiocb *          iocb;           /* controlling i/o request */
  74         struct inode *          inode;          /* target file of i/o */
  75
  76         /* completion state */
  77         atomic_t                io_count;       /* i/os we're waiting for */
  78         spinlock_t              lock;           /* protect completion state */
  79
  80         loff_t                  io_start;       /* Start offset for I/O */
  81         ssize_t                 count,          /* bytes actually processed */
  82                                 max_count,      /* max expected count */
  83                                 bytes_left,     /* bytes left to be sent */
  84                                 error;          /* any reported error */
  85         struct completion       completion;     /* wait for i/o completion */
  86
  87         /* commit state */
  88         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  89         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  90         struct work_struct      work;
  91         int                     flags;
  92         /* for write */
  93 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  94 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  95         /* for read */
  96 #define NFS_ODIRECT_SHOULD_DIRTY        (3)     /* dirty user-space page after read */
  97         struct nfs_writeverf    verf;           /* unstable write verifier */
  98 };
  99
 100 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 101 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 102 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
 103 static void nfs_direct_write_schedule_work(struct work_struct *work);
 104
 105 static inline void get_dreq(struct nfs_direct_req *dreq)
 106 {
 107         atomic_inc(&dreq->io_count);
 108 }
 109
 110 static inline int put_dreq(struct nfs_direct_req *dreq)
 111 {
 112         return atomic_dec_and_test(&dreq->io_count);
 113 }
 114
 115 static void
 116 nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
 117                             const struct nfs_pgio_header *hdr,
 118                             ssize_t dreq_len)
 119 {
 120         if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
 121               test_bit(NFS_IOHDR_EOF, &hdr->flags)))
 122                 return;
 123         if (dreq->max_count >= dreq_len) {
 124                 dreq->max_count = dreq_len;
 125                 if (dreq->count > dreq_len)
 126                         dreq->count = dreq_len;
 127
 128                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 129                         dreq->error = hdr->error;
 130                 else /* Clear outstanding error if this is EOF */
 131                         dreq->error = 0;
 132         }
 133 }
 134
 135 static void
 136 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 137                        const struct nfs_pgio_header *hdr)
 138 {
 139         loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 140         ssize_t dreq_len = 0;
 141
 142         if (hdr_end > dreq->io_start)
 143                 dreq_len = hdr_end - dreq->io_start;
 144
 145         nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 146
 147         if (dreq_len > dreq->max_count)
 148                 dreq_len = dreq->max_count;
 149
 150         if (dreq->count < dreq_len)
 151                 dreq->count = dreq_len;
 152 }
 153
 154 /*
 155  * nfs_direct_select_verf - select the right verifier
 156  * @dreq - direct request possibly spanning multiple servers
 157  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 158  * @commit_idx - commit bucket index for the DS
 159  *
 160  * returns the correct verifier to use given the role of the server
 161  */
 162 static struct nfs_writeverf *
 163 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 164                        struct nfs_client *ds_clp,
 165                        int commit_idx)
 166 {
 167         struct nfs_writeverf *verfp = &dreq->verf;
 168
 169 #ifdef CONFIG_NFS_V4_1
 170         /*
 171          * pNFS is in use, use the DS verf except commit_through_mds is set
 172          * for layout segment where nbuckets is zero.
 173          */
 174         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 175                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 176                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 177                 else
 178                         WARN_ON_ONCE(1);
 179         }
 180 #endif
 181         return verfp;
 182 }
 183
 184
 185 /*
 186  * nfs_direct_set_hdr_verf - set the write/commit verifier
 187  * @dreq - direct request possibly spanning multiple servers
 188  * @hdr - pageio header to validate against previously seen verfs
 189  *
 190  * Set the server's (MDS or DS) "seen" verifier
 191  */
 192 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 193                                     struct nfs_pgio_header *hdr)
 194 {
 195         struct nfs_writeverf *verfp;
 196
 197         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 198         WARN_ON_ONCE(verfp->committed >= 0);
 199         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 200         WARN_ON_ONCE(verfp->committed < 0);
 201 }
 202
 203 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
 204                 const struct nfs_writeverf *v2)
 205 {
 206         return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
 207 }
 208
 209 /*
 210  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 211  * @dreq - direct request possibly spanning multiple servers
 212  * @hdr - pageio header to validate against previously seen verf
 213  *
 214  * set the server's "seen" verf if not initialized.
 215  * returns result of comparison between @hdr->verf and the "seen"
 216  * verf of the server used by @hdr (DS or MDS)
 217  */
 218 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 219                                           struct nfs_pgio_header *hdr)
 220 {
 221         struct nfs_writeverf *verfp;
 222
 223         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 224         if (verfp->committed < 0) {
 225                 nfs_direct_set_hdr_verf(dreq, hdr);
 226                 return 0;
 227         }
 228         return nfs_direct_cmp_verf(verfp, &hdr->verf);
 229 }
 230
 231 /*
 232  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 233  * @dreq - direct request possibly spanning multiple servers
 234  * @data - commit data to validate against previously seen verf
 235  *
 236  * returns result of comparison between @data->verf and the verf of
 237  * the server used by @data (DS or MDS)
 238  */
 239 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 240                                            struct nfs_commit_data *data)
 241 {
 242         struct nfs_writeverf *verfp;
 243
 244         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 245                                          data->ds_commit_index);
 246
 247         /* verifier not set so always fail */
 248         if (verfp->committed < 0 || data->res.verf->committed <= NFS_UNSTABLE)
 249                 return 1;
 250
 251         return nfs_direct_cmp_verf(verfp, data->res.verf);
 252 }
 253
 254 /**
 255  * nfs_direct_IO - NFS address space operation for direct I/O
 256  * @iocb: target I/O control block
 257  * @iter: I/O buffer
 258  *
 259  * The presence of this routine in the address space ops vector means
 260  * the NFS client supports direct I/O. However, for most direct IO, we
 261  * shunt off direct read and write requests before the VFS gets them,
 262  * so this method is only ever called for swap.
 263  */
 264 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 265 {
 266         struct inode *inode = iocb->ki_filp->f_mapping->host;
 267
 268         /* we only support swap file calling nfs_direct_IO */
 269         if (!IS_SWAPFILE(inode))
 270                 return 0;
 271
 272         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 273
 274         if (iov_iter_rw(iter) == READ)
 275                 return nfs_file_direct_read(iocb, iter, true);
 276         return nfs_file_direct_write(iocb, iter, true);
 277 }
 278
 279 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 280 {
 281         unsigned int i;
 282         for (i = 0; i < npages; i++)
 283                 put_page(pages[i]);
 284 }
 285
 286 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 287                               struct nfs_direct_req *dreq)
 288 {
 289         cinfo->inode = dreq->inode;
 290         cinfo->mds = &dreq->mds_cinfo;
 291         cinfo->ds = &dreq->ds_cinfo;
 292         cinfo->dreq = dreq;
 293         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 294 }
 295
 296 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 297 {
 298         struct nfs_direct_req *dreq;
 299
 300         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 301         if (!dreq)
 302                 return NULL;
 303
 304         kref_init(&dreq->kref);
 305         kref_get(&dreq->kref);
 306         init_completion(&dreq->completion);
 307         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 308         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 309         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 310         spin_lock_init(&dreq->lock);
 311
 312         return dreq;
 313 }
 314
 315 static void nfs_direct_req_free(struct kref *kref)
 316 {
 317         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 318
 319         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 320         if (dreq->l_ctx != NULL)
 321                 nfs_put_lock_context(dreq->l_ctx);
 322         if (dreq->ctx != NULL)
 323                 put_nfs_open_context(dreq->ctx);
 324         kmem_cache_free(nfs_direct_cachep, dreq);
 325 }
 326
 327 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 328 {
 329         kref_put(&dreq->kref, nfs_direct_req_free);
 330 }
 331
 332 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 333 {
 334         return dreq->bytes_left;
 335 }
 336 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 337
 338 /*
 339  * Collects and returns the final error value/byte-count.
 340  */
 341 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 342 {
 343         ssize_t result = -EIOCBQUEUED;
 344
 345         /* Async requests don't wait here */
 346         if (dreq->iocb)
 347                 goto out;
 348
 349         result = wait_for_completion_killable(&dreq->completion);
 350
 351         if (!result) {
 352                 result = dreq->count;
 353                 WARN_ON_ONCE(dreq->count < 0);
 354         }
 355         if (!result)
 356                 result = dreq->error;
 357
 358 out:
 359         return (ssize_t) result;
 360 }
 361
 362 /*
 363  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 364  * the iocb is still valid here if this is a synchronous request.
 365  */
 366 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 367 {
 368         struct inode *inode = dreq->inode;
 369
 370         inode_dio_end(inode);
 371
 372         if (dreq->iocb) {
 373                 long res = (long) dreq->error;
 374                 if (dreq->count != 0) {
 375                         res = (long) dreq->count;
 376                         WARN_ON_ONCE(dreq->count < 0);
 377                 }
 378                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 379         }
 380
 381         complete(&dreq->completion);
 382
 383         nfs_direct_req_release(dreq);
 384 }
 385
 386 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 387 {
 388         unsigned long bytes = 0;
 389         struct nfs_direct_req *dreq = hdr->dreq;
 390
 391         spin_lock(&dreq->lock);
 392         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 393                 spin_unlock(&dreq->lock);
 394                 goto out_put;
 395         }
 396
 397         nfs_direct_count_bytes(dreq, hdr);
 398         spin_unlock(&dreq->lock);
 399
 400         while (!list_empty(&hdr->pages)) {
 401                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 402                 struct page *page = req->wb_page;
 403
 404                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
 405                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 406                         set_page_dirty(page);
 407                 bytes += req->wb_bytes;
 408                 nfs_list_remove_request(req);
 409                 nfs_release_request(req);
 410         }
 411 out_put:
 412         if (put_dreq(dreq))
 413                 nfs_direct_complete(dreq);
 414         hdr->release(hdr);
 415 }
 416
 417 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 418 {
 419         struct nfs_page *req;
 420
 421         while (!list_empty(head)) {
 422                 req = nfs_list_entry(head->next);
 423                 nfs_list_remove_request(req);
 424                 nfs_release_request(req);
 425         }
 426 }
 427
 428 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 429 {
 430         get_dreq(hdr->dreq);
 431 }
 432
 433 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 434         .error_cleanup = nfs_read_sync_pgio_error,
 435         .init_hdr = nfs_direct_pgio_init,
 436         .completion = nfs_direct_read_completion,
 437 };
 438
 439 /*
 440  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 441  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 442  * bail and stop sending more reads.  Read length accounting is
 443  * handled automatically by nfs_direct_read_result().  Otherwise, if
 444  * no requests have been sent, just return an error.
 445  */
 446
 447 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 448                                               struct iov_iter *iter,
 449                                               loff_t pos)
 450 {
 451         struct nfs_pageio_descriptor desc;
 452         struct inode *inode = dreq->inode;
 453         ssize_t result = -EINVAL;
 454         size_t requested_bytes = 0;
 455         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 456
 457         nfs_pageio_init_read(&desc, dreq->inode, false,
 458                              &nfs_direct_read_completion_ops);
 459         get_dreq(dreq);
 460         desc.pg_dreq = dreq;
 461         inode_dio_begin(inode);
 462
 463         while (iov_iter_count(iter)) {
 464                 struct page **pagevec;
 465                 size_t bytes;
 466                 size_t pgbase;
 467                 unsigned npages, i;
 468
 469                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 470                                                   rsize, &pgbase);
 471                 if (result < 0)
 472                         break;
 473
 474                 bytes = result;
 475                 iov_iter_advance(iter, bytes);
 476                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 477                 for (i = 0; i < npages; i++) {
 478                         struct nfs_page *req;
 479                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 480                         /* XXX do we need to do the eof zeroing found in async_filler? */
 481                         req = nfs_create_request(dreq->ctx, pagevec[i],
 482                                                  pgbase, req_len);
 483                         if (IS_ERR(req)) {
 484                                 result = PTR_ERR(req);
 485                                 break;
 486                         }
 487                         req->wb_index = pos >> PAGE_SHIFT;
 488                         req->wb_offset = pos & ~PAGE_MASK;
 489                         if (!nfs_pageio_add_request(&desc, req)) {
 490                                 result = desc.pg_error;
 491                                 nfs_release_request(req);
 492                                 break;
 493                         }
 494                         pgbase = 0;
 495                         bytes -= req_len;
 496                         requested_bytes += req_len;
 497                         pos += req_len;
 498                         dreq->bytes_left -= req_len;
 499                 }
 500                 nfs_direct_release_pages(pagevec, npages);
 501                 kvfree(pagevec);
 502                 if (result < 0)
 503                         break;
 504         }
 505
 506         nfs_pageio_complete(&desc);
 507
 508         /*
 509          * If no bytes were started, return the error, and let the
 510          * generic layer handle the completion.
 511          */
 512         if (requested_bytes == 0) {
 513                 inode_dio_end(inode);
 514                 nfs_direct_req_release(dreq);
 515                 return result < 0 ? result : -EIO;
 516         }
 517
 518         if (put_dreq(dreq))
 519                 nfs_direct_complete(dreq);
 520         return requested_bytes;
 521 }
 522
 523 /**
 524  * nfs_file_direct_read - file direct read operation for NFS files
 525  * @iocb: target I/O control block
 526  * @iter: vector of user buffers into which to read data
 527  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 528  *
 529  * We use this function for direct reads instead of calling
 530  * generic_file_aio_read() in order to avoid gfar's check to see if
 531  * the request starts before the end of the file.  For that check
 532  * to work, we must generate a GETATTR before each direct read, and
 533  * even then there is a window between the GETATTR and the subsequent
 534  * READ where the file size could change.  Our preference is simply
 535  * to do all reads the application wants, and the server will take
 536  * care of managing the end of file boundary.
 537  *
 538  * This function also eliminates unnecessarily updating the file's
 539  * atime locally, as the NFS server sets the file's atime, and this
 540  * client must read the updated atime from the server back into its
 541  * cache.
 542  */
 543 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 544                              bool swap)
 545 {
 546         struct file *file = iocb->ki_filp;
 547         struct address_space *mapping = file->f_mapping;
 548         struct inode *inode = mapping->host;
 549         struct nfs_direct_req *dreq;
 550         struct nfs_lock_context *l_ctx;
 551         ssize_t result = -EINVAL, requested;
 552         size_t count = iov_iter_count(iter);
 553         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 554
 555         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 556                 file, count, (long long) iocb->ki_pos);
 557
 558         result = 0;
 559         if (!count)
 560                 goto out;
 561
 562         task_io_account_read(count);
 563
 564         result = -ENOMEM;
 565         dreq = nfs_direct_req_alloc();
 566         if (dreq == NULL)
 567                 goto out;
 568
 569         dreq->inode = inode;
 570         dreq->bytes_left = dreq->max_count = count;
 571         dreq->io_start = iocb->ki_pos;
 572         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 573         l_ctx = nfs_get_lock_context(dreq->ctx);
 574         if (IS_ERR(l_ctx)) {
 575                 result = PTR_ERR(l_ctx);
 576                 nfs_direct_req_release(dreq);
 577                 goto out_release;
 578         }
 579         dreq->l_ctx = l_ctx;
 580         if (!is_sync_kiocb(iocb))
 581                 dreq->iocb = iocb;
 582
 583         if (iter_is_iovec(iter))
 584                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 585
 586         if (!swap)
 587                 nfs_start_io_direct(inode);
 588
 589         NFS_I(inode)->read_io += count;
 590         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 591
 592         if (!swap)
 593                 nfs_end_io_direct(inode);
 594
 595         if (requested > 0) {
 596                 result = nfs_direct_wait(dreq);
 597                 if (result > 0) {
 598                         requested -= result;
 599                         iocb->ki_pos += result;
 600                 }
 601                 iov_iter_revert(iter, requested);
 602         } else {
 603                 result = requested;
 604         }
 605
 606 out_release:
 607         nfs_direct_req_release(dreq);
 608 out:
 609         return result;
 610 }
 611
 612 static void
 613 nfs_direct_write_scan_commit_list(struct inode *inode,
 614                                   struct list_head *list,
 615                                   struct nfs_commit_info *cinfo)
 616 {
 617         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 618 #ifdef CONFIG_NFS_V4_1
 619         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 620                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 621 #endif
 622         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 623         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 624 }
 625
 626 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 627 {
 628         struct nfs_pageio_descriptor desc;
 629         struct nfs_page *req, *tmp;
 630         LIST_HEAD(reqs);
 631         struct nfs_commit_info cinfo;
 632         LIST_HEAD(failed);
 633
 634         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 635         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 636
 637         dreq->count = 0;
 638         dreq->max_count = 0;
 639         list_for_each_entry(req, &reqs, wb_list)
 640                 dreq->max_count += req->wb_bytes;
 641         dreq->verf.committed = NFS_INVALID_STABLE_HOW;
 642         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 643         get_dreq(dreq);
 644
 645         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 646                               &nfs_direct_write_completion_ops);
 647         desc.pg_dreq = dreq;
 648
 649         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 650                 /* Bump the transmission count */
 651                 req->wb_nio++;
 652                 if (!nfs_pageio_add_request(&desc, req)) {
 653                         nfs_list_move_request(req, &failed);
 654                         spin_lock(&cinfo.inode->i_lock);
 655                         dreq->flags = 0;
 656                         if (desc.pg_error < 0)
 657                                 dreq->error = desc.pg_error;
 658                         else
 659                                 dreq->error = -EIO;
 660                         spin_unlock(&cinfo.inode->i_lock);
 661                 }
 662                 nfs_release_request(req);
 663         }
 664         nfs_pageio_complete(&desc);
 665
 666         while (!list_empty(&failed)) {
 667                 req = nfs_list_entry(failed.next);
 668                 nfs_list_remove_request(req);
 669                 nfs_unlock_and_release_request(req);
 670         }
 671
 672         if (put_dreq(dreq))
 673                 nfs_direct_write_complete(dreq);
 674 }
 675
 676 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 677 {
 678         struct nfs_direct_req *dreq = data->dreq;
 679         struct nfs_commit_info cinfo;
 680         struct nfs_page *req;
 681         int status = data->task.tk_status;
 682
 683         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 684         if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
 685                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 686
 687         while (!list_empty(&data->pages)) {
 688                 req = nfs_list_entry(data->pages.next);
 689                 nfs_list_remove_request(req);
 690                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 691                         /*
 692                          * Despite the reboot, the write was successful,
 693                          * so reset wb_nio.
 694                          */
 695                         req->wb_nio = 0;
 696                         /* Note the rewrite will go through mds */
 697                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 698                 } else
 699                         nfs_release_request(req);
 700                 nfs_unlock_and_release_request(req);
 701         }
 702
 703         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 704                 nfs_direct_write_complete(dreq);
 705 }
 706
 707 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 708                 struct nfs_page *req)
 709 {
 710         struct nfs_direct_req *dreq = cinfo->dreq;
 711
 712         spin_lock(&dreq->lock);
 713         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 714         spin_unlock(&dreq->lock);
 715         nfs_mark_request_commit(req, NULL, cinfo, 0);
 716 }
 717
 718 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 719         .completion = nfs_direct_commit_complete,
 720         .resched_write = nfs_direct_resched_write,
 721 };
 722
 723 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 724 {
 725         int res;
 726         struct nfs_commit_info cinfo;
 727         LIST_HEAD(mds_list);
 728
 729         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 730         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 731         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 732         if (res < 0) /* res == -ENOMEM */
 733                 nfs_direct_write_reschedule(dreq);
 734 }
 735
 736 static void nfs_direct_write_schedule_work(struct work_struct *work)
 737 {
 738         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 739         int flags = dreq->flags;
 740
 741         dreq->flags = 0;
 742         switch (flags) {
 743                 case NFS_ODIRECT_DO_COMMIT:
 744                         nfs_direct_commit_schedule(dreq);
 745                         break;
 746                 case NFS_ODIRECT_RESCHED_WRITES:
 747                         nfs_direct_write_reschedule(dreq);
 748                         break;
 749                 default:
 750                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 751                         nfs_direct_complete(dreq);
 752         }
 753 }
 754
 755 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 756 {
 757         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 758 }
 759
 760 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 761 {
 762         struct nfs_direct_req *dreq = hdr->dreq;
 763         struct nfs_commit_info cinfo;
 764         bool request_commit = false;
 765         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 766
 767         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 768
 769         spin_lock(&dreq->lock);
 770         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 771                 spin_unlock(&dreq->lock);
 772                 goto out_put;
 773         }
 774
 775         nfs_direct_count_bytes(dreq, hdr);
 776         if (hdr->good_bytes != 0) {
 777                 if (nfs_write_need_commit(hdr)) {
 778                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 779                                 request_commit = true;
 780                         else if (dreq->flags == 0) {
 781                                 nfs_direct_set_hdr_verf(dreq, hdr);
 782                                 request_commit = true;
 783                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 784                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 785                                 request_commit = true;
 786                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 787                                         dreq->flags =
 788                                                 NFS_ODIRECT_RESCHED_WRITES;
 789                         }
 790                 }
 791         }
 792         spin_unlock(&dreq->lock);
 793
 794         while (!list_empty(&hdr->pages)) {
 795
 796                 req = nfs_list_entry(hdr->pages.next);
 797                 nfs_list_remove_request(req);
 798                 if (request_commit) {
 799                         kref_get(&req->wb_kref);
 800                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 801                                 hdr->ds_commit_idx);
 802                 }
 803                 nfs_unlock_and_release_request(req);
 804         }
 805
 806 out_put:
 807         if (put_dreq(dreq))
 808                 nfs_direct_write_complete(dreq);
 809         hdr->release(hdr);
 810 }
 811
 812 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 813 {
 814         struct nfs_page *req;
 815
 816         while (!list_empty(head)) {
 817                 req = nfs_list_entry(head->next);
 818                 nfs_list_remove_request(req);
 819                 nfs_unlock_and_release_request(req);
 820         }
 821 }
 822
 823 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 824 {
 825         struct nfs_direct_req *dreq = hdr->dreq;
 826
 827         spin_lock(&dreq->lock);
 828         if (dreq->error == 0) {
 829                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 830                 /* fake unstable write to let common nfs resend pages */
 831                 hdr->verf.committed = NFS_UNSTABLE;
 832                 hdr->good_bytes = hdr->args.count;
 833         }
 834         spin_unlock(&dreq->lock);
 835 }
 836
 837 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 838         .error_cleanup = nfs_write_sync_pgio_error,
 839         .init_hdr = nfs_direct_pgio_init,
 840         .completion = nfs_direct_write_completion,
 841         .reschedule_io = nfs_direct_write_reschedule_io,
 842 };
 843
 844
 845 /*
 846  * NB: Return the value of the first error return code.  Subsequent
 847  *     errors after the first one are ignored.
 848  */
 849 /*
 850  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 851  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 852  * bail and stop sending more writes.  Write length accounting is
 853  * handled automatically by nfs_direct_write_result().  Otherwise, if
 854  * no requests have been sent, just return an error.
 855  */
 856 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 857                                                struct iov_iter *iter,
 858                                                loff_t pos, int ioflags)
 859 {
 860         struct nfs_pageio_descriptor desc;
 861         struct inode *inode = dreq->inode;
 862         ssize_t result = 0;
 863         size_t requested_bytes = 0;
 864         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 865
 866         nfs_pageio_init_write(&desc, inode, ioflags, false,
 867                               &nfs_direct_write_completion_ops);
 868         desc.pg_dreq = dreq;
 869         get_dreq(dreq);
 870         inode_dio_begin(inode);
 871
 872         NFS_I(inode)->write_io += iov_iter_count(iter);
 873         while (iov_iter_count(iter)) {
 874                 struct page **pagevec;
 875                 size_t bytes;
 876                 size_t pgbase;
 877                 unsigned npages, i;
 878
 879                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 880                                                   wsize, &pgbase);
 881                 if (result < 0)
 882                         break;
 883
 884                 bytes = result;
 885                 iov_iter_advance(iter, bytes);
 886                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 887                 for (i = 0; i < npages; i++) {
 888                         struct nfs_page *req;
 889                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 890
 891                         req = nfs_create_request(dreq->ctx, pagevec[i],
 892                                                  pgbase, req_len);
 893                         if (IS_ERR(req)) {
 894                                 result = PTR_ERR(req);
 895                                 break;
 896                         }
 897
 898                         if (desc.pg_error < 0) {
 899                                 nfs_free_request(req);
 900                                 result = desc.pg_error;
 901                                 break;
 902                         }
 903
 904                         nfs_lock_request(req);
 905                         req->wb_index = pos >> PAGE_SHIFT;
 906                         req->wb_offset = pos & ~PAGE_MASK;
 907                         if (!nfs_pageio_add_request(&desc, req)) {
 908                                 result = desc.pg_error;
 909                                 nfs_unlock_and_release_request(req);
 910                                 break;
 911                         }
 912                         pgbase = 0;
 913                         bytes -= req_len;
 914                         requested_bytes += req_len;
 915                         pos += req_len;
 916                         dreq->bytes_left -= req_len;
 917                 }
 918                 nfs_direct_release_pages(pagevec, npages);
 919                 kvfree(pagevec);
 920                 if (result < 0)
 921                         break;
 922         }
 923         nfs_pageio_complete(&desc);
 924
 925         /*
 926          * If no bytes were started, return the error, and let the
 927          * generic layer handle the completion.
 928          */
 929         if (requested_bytes == 0) {
 930                 inode_dio_end(inode);
 931                 nfs_direct_req_release(dreq);
 932                 return result < 0 ? result : -EIO;
 933         }
 934
 935         if (put_dreq(dreq))
 936                 nfs_direct_write_complete(dreq);
 937         return requested_bytes;
 938 }
 939
 940 /**
 941  * nfs_file_direct_write - file direct write operation for NFS files
 942  * @iocb: target I/O control block
 943  * @iter: vector of user buffers from which to write data
 944  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 945  *
 946  * We use this function for direct writes instead of calling
 947  * generic_file_aio_write() in order to avoid taking the inode
 948  * semaphore and updating the i_size.  The NFS server will set
 949  * the new i_size and this client must read the updated size
 950  * back into its cache.  We let the server do generic write
 951  * parameter checking and report problems.
 952  *
 953  * We eliminate local atime updates, see direct read above.
 954  *
 955  * We avoid unnecessary page cache invalidations for normal cached
 956  * readers of this file.
 957  *
 958  * Note that O_APPEND is not supported for NFS direct writes, as there
 959  * is no atomic O_APPEND write facility in the NFS protocol.
 960  */
 961 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 962                               bool swap)
 963 {
 964         ssize_t result = -EINVAL, requested;
 965         size_t count;
 966         struct file *file = iocb->ki_filp;
 967         struct address_space *mapping = file->f_mapping;
 968         struct inode *inode = mapping->host;
 969         struct nfs_direct_req *dreq;
 970         struct nfs_lock_context *l_ctx;
 971         loff_t pos, end;
 972
 973         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 974                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 975
 976         if (swap)
 977                 /* bypass generic checks */
 978                 result =  iov_iter_count(iter);
 979         else
 980                 result = generic_write_checks(iocb, iter);
 981         if (result <= 0)
 982                 return result;
 983         count = result;
 984         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 985
 986         pos = iocb->ki_pos;
 987         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 988
 989         task_io_account_write(count);
 990
 991         result = -ENOMEM;
 992         dreq = nfs_direct_req_alloc();
 993         if (!dreq)
 994                 goto out;
 995
 996         dreq->inode = inode;
 997         dreq->bytes_left = dreq->max_count = count;
 998         dreq->io_start = pos;
 999         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1000         l_ctx = nfs_get_lock_context(dreq->ctx);
1001         if (IS_ERR(l_ctx)) {
1002                 result = PTR_ERR(l_ctx);
1003                 nfs_direct_req_release(dreq);
1004                 goto out_release;
1005         }
1006         dreq->l_ctx = l_ctx;
1007         if (!is_sync_kiocb(iocb))
1008                 dreq->iocb = iocb;
1009
1010         if (swap) {
1011                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1012                                                             FLUSH_STABLE);
1013         } else {
1014                 nfs_start_io_direct(inode);
1015
1016                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1017                                                             FLUSH_COND_STABLE);
1018
1019                 if (mapping->nrpages) {
1020                         invalidate_inode_pages2_range(mapping,
1021                                                       pos >> PAGE_SHIFT, end);
1022                 }
1023
1024                 nfs_end_io_direct(inode);
1025         }
1026
1027         if (requested > 0) {
1028                 result = nfs_direct_wait(dreq);
1029                 if (result > 0) {
1030                         requested -= result;
1031                         iocb->ki_pos = pos + result;
1032                         /* XXX: should check the generic_write_sync retval */
1033                         generic_write_sync(iocb, result);
1034                 }
1035                 iov_iter_revert(iter, requested);
1036         } else {
1037                 result = requested;
1038         }
1039 out_release:
1040         nfs_direct_req_release(dreq);
1041 out:
1042         return result;
1043 }
1044
1045 /**
1046  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1047  *
1048  */
1049 int __init nfs_init_directcache(void)
1050 {
1051         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1052                                                 sizeof(struct nfs_direct_req),
1053                                                 0, (SLAB_RECLAIM_ACCOUNT|
1054                                                         SLAB_MEM_SPREAD),
1055                                                 NULL);
1056         if (nfs_direct_cachep == NULL)
1057                 return -ENOMEM;
1058
1059         return 0;
1060 }
1061
1062 /**
1063  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1064  *
1065  */
1066 void nfs_destroy_directcache(void)
1067 {
1068         kmem_cache_destroy(nfs_direct_cachep);
1069 }