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