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_req {
  70         struct kref             kref;           /* release manager */
  71
  72         /* I/O parameters */
  73         struct nfs_open_context *ctx;           /* file open context info */
  74         struct nfs_lock_context *l_ctx;         /* Lock context info */
  75         struct kiocb *          iocb;           /* controlling i/o request */
  76         struct inode *          inode;          /* target file of i/o */
  77
  78         /* completion state */
  79         atomic_t                io_count;       /* i/os we're waiting for */
  80         spinlock_t              lock;           /* protect completion state */
  81         ssize_t                 count,          /* bytes actually processed */
  82                                 bytes_left,     /* bytes left to be sent */
  83                                 error;          /* any reported error */
  84         struct completion       completion;     /* wait for i/o completion */
  85
  86         /* commit state */
  87         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  88         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  89         struct work_struct      work;
  90         int                     flags;
  91 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  92 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  93 #define NFS_ODIRECT_MARK_DIRTY          (4)     /* mark read pages dirty */
  94         struct nfs_writeverf    verf;           /* unstable write verifier */
  95 };
  96
  97 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  98 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  99 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
 100 static void nfs_direct_write_schedule_work(struct work_struct *work);
 101
 102 static inline void get_dreq(struct nfs_direct_req *dreq)
 103 {
 104         atomic_inc(&dreq->io_count);
 105 }
 106
 107 static inline int put_dreq(struct nfs_direct_req *dreq)
 108 {
 109         return atomic_dec_and_test(&dreq->io_count);
 110 }
 111
 112 /**
 113  * nfs_direct_IO - NFS address space operation for direct I/O
 114  * @rw: direction (read or write)
 115  * @iocb: target I/O control block
 116  * @iter: array of vectors that define I/O buffer
 117  * @pos: offset in file to begin the operation
 118  * @nr_segs: size of iovec array
 119  *
 120  * The presence of this routine in the address space ops vector means
 121  * the NFS client supports direct I/O. However, we shunt off direct
 122  * read and write requests before the VFS gets them, so this method
 123  * should never be called.
 124  */
 125 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter,
 126                       loff_t pos)
 127 {
 128         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 129                         iocb->ki_filp, (long long) pos, iter->nr_segs);
 130
 131         return -EINVAL;
 132 }
 133
 134 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 135 {
 136         unsigned int i;
 137         for (i = 0; i < npages; i++)
 138                 page_cache_release(pages[i]);
 139 }
 140
 141 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 142                               struct nfs_direct_req *dreq)
 143 {
 144         cinfo->lock = &dreq->lock;
 145         cinfo->mds = &dreq->mds_cinfo;
 146         cinfo->ds = &dreq->ds_cinfo;
 147         cinfo->dreq = dreq;
 148         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 149 }
 150
 151 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 152 {
 153         struct nfs_direct_req *dreq;
 154
 155         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 156         if (!dreq)
 157                 return NULL;
 158
 159         kref_init(&dreq->kref);
 160         kref_get(&dreq->kref);
 161         init_completion(&dreq->completion);
 162         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 163         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 164         spin_lock_init(&dreq->lock);
 165
 166         return dreq;
 167 }
 168
 169 static void nfs_direct_req_free(struct kref *kref)
 170 {
 171         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 172
 173         if (dreq->l_ctx != NULL)
 174                 nfs_put_lock_context(dreq->l_ctx);
 175         if (dreq->ctx != NULL)
 176                 put_nfs_open_context(dreq->ctx);
 177         kmem_cache_free(nfs_direct_cachep, dreq);
 178 }
 179
 180 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 181 {
 182         kref_put(&dreq->kref, nfs_direct_req_free);
 183 }
 184
 185 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 186 {
 187         return dreq->bytes_left;
 188 }
 189 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 190
 191 /*
 192  * Collects and returns the final error value/byte-count.
 193  */
 194 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 195 {
 196         ssize_t result = -EIOCBQUEUED;
 197
 198         /* Async requests don't wait here */
 199         if (dreq->iocb)
 200                 goto out;
 201
 202         result = wait_for_completion_killable(&dreq->completion);
 203
 204         if (!result)
 205                 result = dreq->error;
 206         if (!result)
 207                 result = dreq->count;
 208
 209 out:
 210         return (ssize_t) result;
 211 }
 212
 213 /*
 214  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 215  * the iocb is still valid here if this is a synchronous request.
 216  */
 217 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 218 {
 219         if (dreq->iocb) {
 220                 long res = (long) dreq->error;
 221                 if (!res)
 222                         res = (long) dreq->count;
 223                 aio_complete(dreq->iocb, res, 0);
 224         }
 225         complete_all(&dreq->completion);
 226
 227         nfs_direct_req_release(dreq);
 228 }
 229
 230 static void nfs_direct_readpage_release(struct nfs_page *req)
 231 {
 232         dprintk("NFS: direct read done (%s/%lld %d@%lld)\n",
 233                 req->wb_context->dentry->d_inode->i_sb->s_id,
 234                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 235                 req->wb_bytes,
 236                 (long long)req_offset(req));
 237         nfs_release_request(req);
 238 }
 239
 240 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 241 {
 242         unsigned long bytes = 0;
 243         struct nfs_direct_req *dreq = hdr->dreq;
 244
 245         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 246                 goto out_put;
 247
 248         spin_lock(&dreq->lock);
 249         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 250                 dreq->error = hdr->error;
 251         else
 252                 dreq->count += hdr->good_bytes;
 253         spin_unlock(&dreq->lock);
 254
 255         while (!list_empty(&hdr->pages)) {
 256                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 257                 struct page *page = req->wb_page;
 258
 259                 if ((dreq->flags & NFS_ODIRECT_MARK_DIRTY) &&
 260                     !PageCompound(page) && bytes < hdr->good_bytes)
 261                         set_page_dirty(page);
 262                 bytes += req->wb_bytes;
 263                 nfs_list_remove_request(req);
 264                 nfs_direct_readpage_release(req);
 265         }
 266 out_put:
 267         if (put_dreq(dreq))
 268                 nfs_direct_complete(dreq);
 269         hdr->release(hdr);
 270 }
 271
 272 static void nfs_read_sync_pgio_error(struct list_head *head)
 273 {
 274         struct nfs_page *req;
 275
 276         while (!list_empty(head)) {
 277                 req = nfs_list_entry(head->next);
 278                 nfs_list_remove_request(req);
 279                 nfs_release_request(req);
 280         }
 281 }
 282
 283 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 284 {
 285         get_dreq(hdr->dreq);
 286 }
 287
 288 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 289         .error_cleanup = nfs_read_sync_pgio_error,
 290         .init_hdr = nfs_direct_pgio_init,
 291         .completion = nfs_direct_read_completion,
 292 };
 293
 294 /*
 295  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 296  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 297  * bail and stop sending more reads.  Read length accounting is
 298  * handled automatically by nfs_direct_read_result().  Otherwise, if
 299  * no requests have been sent, just return an error.
 300  */
 301 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
 302                                                 const struct iovec *iov,
 303                                                 loff_t pos)
 304 {
 305         struct nfs_direct_req *dreq = desc->pg_dreq;
 306         struct nfs_open_context *ctx = dreq->ctx;
 307         struct inode *inode = ctx->dentry->d_inode;
 308         unsigned long user_addr = (unsigned long)iov->iov_base;
 309         size_t count = iov->iov_len;
 310         size_t rsize = NFS_SERVER(inode)->rsize;
 311         unsigned int pgbase;
 312         int result;
 313         ssize_t started = 0;
 314         struct page **pagevec = NULL;
 315         unsigned int npages;
 316
 317         do {
 318                 size_t bytes;
 319                 int i;
 320
 321                 pgbase = user_addr & ~PAGE_MASK;
 322                 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
 323
 324                 result = -ENOMEM;
 325                 npages = nfs_page_array_len(pgbase, bytes);
 326                 if (!pagevec)
 327                         pagevec = kmalloc(npages * sizeof(struct page *),
 328                                           GFP_KERNEL);
 329                 if (!pagevec)
 330                         break;
 331                 down_read(&current->mm->mmap_sem);
 332                 result = get_user_pages(current, current->mm, user_addr,
 333                                         npages, 1, 0, pagevec, NULL);
 334                 up_read(&current->mm->mmap_sem);
 335                 if (result < 0)
 336                         break;
 337                 if ((unsigned)result < npages) {
 338                         bytes = result * PAGE_SIZE;
 339                         if (bytes <= pgbase) {
 340                                 nfs_direct_release_pages(pagevec, result);
 341                                 break;
 342                         }
 343                         bytes -= pgbase;
 344                         npages = result;
 345                 }
 346
 347                 for (i = 0; i < npages; i++) {
 348                         struct nfs_page *req;
 349                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 350                         /* XXX do we need to do the eof zeroing found in async_filler? */
 351                         req = nfs_create_request(dreq->ctx, dreq->inode,
 352                                                  pagevec[i],
 353                                                  pgbase, req_len);
 354                         if (IS_ERR(req)) {
 355                                 result = PTR_ERR(req);
 356                                 break;
 357                         }
 358                         req->wb_index = pos >> PAGE_SHIFT;
 359                         req->wb_offset = pos & ~PAGE_MASK;
 360                         if (!nfs_pageio_add_request(desc, req)) {
 361                                 result = desc->pg_error;
 362                                 nfs_release_request(req);
 363                                 break;
 364                         }
 365                         pgbase = 0;
 366                         bytes -= req_len;
 367                         started += req_len;
 368                         user_addr += req_len;
 369                         pos += req_len;
 370                         count -= req_len;
 371                         dreq->bytes_left -= req_len;
 372                 }
 373                 /* The nfs_page now hold references to these pages */
 374                 nfs_direct_release_pages(pagevec, npages);
 375         } while (count != 0 && result >= 0);
 376
 377         kfree(pagevec);
 378
 379         if (started)
 380                 return started;
 381         return result < 0 ? (ssize_t) result : -EFAULT;
 382 }
 383
 384 static ssize_t nfs_direct_do_schedule_read_iovec(
 385                 struct nfs_pageio_descriptor *desc, const struct iovec *iov,
 386                 unsigned long nr_segs, loff_t pos)
 387 {
 388         ssize_t result = -EINVAL;
 389         size_t requested_bytes = 0;
 390         unsigned long seg;
 391
 392         for (seg = 0; seg < nr_segs; seg++) {
 393                 const struct iovec *vec = &iov[seg];
 394                 result = nfs_direct_read_schedule_segment(desc, vec, pos);
 395                 if (result < 0)
 396                         break;
 397                 requested_bytes += result;
 398                 if ((size_t)result < vec->iov_len)
 399                         break;
 400                 pos += vec->iov_len;
 401         }
 402         if (requested_bytes)
 403                 return requested_bytes;
 404
 405         return result < 0 ? result : -EIO;
 406 }
 407
 408 #ifdef CONFIG_BLOCK
 409 static ssize_t nfs_direct_do_schedule_read_bvec(
 410                 struct nfs_pageio_descriptor *desc,
 411                 struct bio_vec *bvec, unsigned long nr_segs, loff_t pos)
 412 {
 413         struct nfs_direct_req *dreq = desc->pg_dreq;
 414         struct nfs_open_context *ctx = dreq->ctx;
 415         struct inode *inode = ctx->dentry->d_inode;
 416         ssize_t result = -EINVAL;
 417         size_t requested_bytes = 0;
 418         unsigned long seg;
 419         struct nfs_page *req;
 420         unsigned int req_len;
 421
 422         for (seg = 0; seg < nr_segs; seg++) {
 423                 result = -EIO;
 424                 req_len = bvec[seg].bv_len;
 425                 req = nfs_create_request(ctx, inode,
 426                                          bvec[seg].bv_page,
 427                                          bvec[seg].bv_offset, req_len);
 428                 if (IS_ERR(req)) {
 429                         result = PTR_ERR(req);
 430                         break;
 431                 }
 432                 req->wb_index = pos >> PAGE_SHIFT;
 433                 req->wb_offset = pos & ~PAGE_MASK;
 434                 if (!nfs_pageio_add_request(desc, req)) {
 435                         result = desc->pg_error;
 436                         nfs_release_request(req);
 437                         break;
 438                 }
 439                 requested_bytes += req_len;
 440                 pos += req_len;
 441         }
 442
 443         if (requested_bytes)
 444                 return requested_bytes;
 445
 446         return result < 0 ? result : -EIO;
 447 }
 448 #endif /* CONFIG_BLOCK */
 449
 450 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq,
 451                                         struct iov_iter *iter, loff_t pos)
 452 {
 453         struct nfs_pageio_descriptor desc;
 454         ssize_t result;
 455
 456         NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
 457                              &nfs_direct_read_completion_ops);
 458         get_dreq(dreq);
 459         desc.pg_dreq = dreq;
 460
 461         if (iov_iter_has_iovec(iter)) {
 462                 result = nfs_direct_do_schedule_read_iovec(&desc,
 463                                 iov_iter_iovec(iter), iter->nr_segs, pos);
 464 #ifdef CONFIG_BLOCK
 465         } else if (iov_iter_has_bvec(iter)) {
 466                 result = nfs_direct_do_schedule_read_bvec(&desc,
 467                                 iov_iter_bvec(iter), iter->nr_segs, pos);
 468 #endif
 469         } else
 470                 BUG();
 471
 472         nfs_pageio_complete(&desc);
 473
 474         /*
 475          * If no bytes were started, return the error, and let the
 476          * generic layer handle the completion.
 477          */
 478         if (result < 0) {
 479                 nfs_direct_req_release(dreq);
 480                 return result;
 481         }
 482
 483         if (put_dreq(dreq))
 484                 nfs_direct_complete(dreq);
 485         return 0;
 486 }
 487
 488 static ssize_t nfs_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 489                                loff_t pos)
 490 {
 491         ssize_t result = -ENOMEM;
 492         struct inode *inode = iocb->ki_filp->f_mapping->host;
 493         struct nfs_direct_req *dreq;
 494         struct nfs_lock_context *l_ctx;
 495
 496         dreq = nfs_direct_req_alloc();
 497         if (dreq == NULL)
 498                 goto out;
 499
 500         dreq->inode = inode;
 501         dreq->bytes_left = iov_iter_count(iter);
 502         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 503         l_ctx = nfs_get_lock_context(dreq->ctx);
 504         if (IS_ERR(l_ctx)) {
 505                 result = PTR_ERR(l_ctx);
 506                 goto out_release;
 507         }
 508         dreq->l_ctx = l_ctx;
 509         if (!is_sync_kiocb(iocb))
 510                 dreq->iocb = iocb;
 511
 512         NFS_I(inode)->read_io += iov_iter_count(iter);
 513         result = nfs_direct_read_schedule(dreq, iter, pos);
 514         if (!result)
 515                 result = nfs_direct_wait(dreq);
 516 out_release:
 517         nfs_direct_req_release(dreq);
 518 out:
 519         return result;
 520 }
 521
 522 static void nfs_inode_dio_write_done(struct inode *inode)
 523 {
 524         nfs_zap_mapping(inode, inode->i_mapping);
 525         inode_dio_done(inode);
 526 }
 527
 528 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 529 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 530 {
 531         struct nfs_pageio_descriptor desc;
 532         struct nfs_page *req, *tmp;
 533         LIST_HEAD(reqs);
 534         struct nfs_commit_info cinfo;
 535         LIST_HEAD(failed);
 536
 537         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 538         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 539         spin_lock(cinfo.lock);
 540         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 541         spin_unlock(cinfo.lock);
 542
 543         dreq->count = 0;
 544         get_dreq(dreq);
 545
 546         NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
 547                               &nfs_direct_write_completion_ops);
 548         desc.pg_dreq = dreq;
 549
 550         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 551                 if (!nfs_pageio_add_request(&desc, req)) {
 552                         nfs_list_remove_request(req);
 553                         nfs_list_add_request(req, &failed);
 554                         spin_lock(cinfo.lock);
 555                         dreq->flags = 0;
 556                         dreq->error = -EIO;
 557                         spin_unlock(cinfo.lock);
 558                 }
 559                 nfs_release_request(req);
 560         }
 561         nfs_pageio_complete(&desc);
 562
 563         while (!list_empty(&failed)) {
 564                 req = nfs_list_entry(failed.next);
 565                 nfs_list_remove_request(req);
 566                 nfs_unlock_and_release_request(req);
 567         }
 568
 569         if (put_dreq(dreq))
 570                 nfs_direct_write_complete(dreq, dreq->inode);
 571 }
 572
 573 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 574 {
 575         struct nfs_direct_req *dreq = data->dreq;
 576         struct nfs_commit_info cinfo;
 577         struct nfs_page *req;
 578         int status = data->task.tk_status;
 579
 580         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 581         if (status < 0) {
 582                 dprintk("NFS: %5u commit failed with error %d.\n",
 583                         data->task.tk_pid, status);
 584                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 585         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 586                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 587                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 588         }
 589
 590         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 591         while (!list_empty(&data->pages)) {
 592                 req = nfs_list_entry(data->pages.next);
 593                 nfs_list_remove_request(req);
 594                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 595                         /* Note the rewrite will go through mds */
 596                         nfs_mark_request_commit(req, NULL, &cinfo);
 597                 } else
 598                         nfs_release_request(req);
 599                 nfs_unlock_and_release_request(req);
 600         }
 601
 602         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 603                 nfs_direct_write_complete(dreq, data->inode);
 604 }
 605
 606 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 607 {
 608         /* There is no lock to clear */
 609 }
 610
 611 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 612         .completion = nfs_direct_commit_complete,
 613         .error_cleanup = nfs_direct_error_cleanup,
 614 };
 615
 616 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 617 {
 618         int res;
 619         struct nfs_commit_info cinfo;
 620         LIST_HEAD(mds_list);
 621
 622         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 623         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 624         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 625         if (res < 0) /* res == -ENOMEM */
 626                 nfs_direct_write_reschedule(dreq);
 627 }
 628
 629 static void nfs_direct_write_schedule_work(struct work_struct *work)
 630 {
 631         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 632         int flags = dreq->flags;
 633
 634         dreq->flags = 0;
 635         switch (flags) {
 636                 case NFS_ODIRECT_DO_COMMIT:
 637                         nfs_direct_commit_schedule(dreq);
 638                         break;
 639                 case NFS_ODIRECT_RESCHED_WRITES:
 640                         nfs_direct_write_reschedule(dreq);
 641                         break;
 642                 default:
 643                         nfs_inode_dio_write_done(dreq->inode);
 644                         nfs_direct_complete(dreq);
 645         }
 646 }
 647
 648 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 649 {
 650         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 651 }
 652
 653 #else
 654 static void nfs_direct_write_schedule_work(struct work_struct *work)
 655 {
 656 }
 657
 658 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 659 {
 660         nfs_inode_dio_write_done(inode);
 661         nfs_direct_complete(dreq);
 662 }
 663 #endif
 664
 665 /*
 666  * NB: Return the value of the first error return code.  Subsequent
 667  *     errors after the first one are ignored.
 668  */
 669 /*
 670  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 671  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 672  * bail and stop sending more writes.  Write length accounting is
 673  * handled automatically by nfs_direct_write_result().  Otherwise, if
 674  * no requests have been sent, just return an error.
 675  */
 676 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
 677                                                  const struct iovec *iov,
 678                                                  loff_t pos)
 679 {
 680         struct nfs_direct_req *dreq = desc->pg_dreq;
 681         struct nfs_open_context *ctx = dreq->ctx;
 682         struct inode *inode = ctx->dentry->d_inode;
 683         unsigned long user_addr = (unsigned long)iov->iov_base;
 684         size_t count = iov->iov_len;
 685         size_t wsize = NFS_SERVER(inode)->wsize;
 686         unsigned int pgbase;
 687         int result;
 688         ssize_t started = 0;
 689         struct page **pagevec = NULL;
 690         unsigned int npages;
 691
 692         do {
 693                 size_t bytes;
 694                 int i;
 695
 696                 pgbase = user_addr & ~PAGE_MASK;
 697                 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
 698
 699                 result = -ENOMEM;
 700                 npages = nfs_page_array_len(pgbase, bytes);
 701                 if (!pagevec)
 702                         pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
 703                 if (!pagevec)
 704                         break;
 705
 706                 down_read(&current->mm->mmap_sem);
 707                 result = get_user_pages(current, current->mm, user_addr,
 708                                         npages, 0, 0, pagevec, NULL);
 709                 up_read(&current->mm->mmap_sem);
 710                 if (result < 0)
 711                         break;
 712
 713                 if ((unsigned)result < npages) {
 714                         bytes = result * PAGE_SIZE;
 715                         if (bytes <= pgbase) {
 716                                 nfs_direct_release_pages(pagevec, result);
 717                                 break;
 718                         }
 719                         bytes -= pgbase;
 720                         npages = result;
 721                 }
 722
 723                 for (i = 0; i < npages; i++) {
 724                         struct nfs_page *req;
 725                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 726
 727                         req = nfs_create_request(dreq->ctx, dreq->inode,
 728                                                  pagevec[i],
 729                                                  pgbase, req_len);
 730                         if (IS_ERR(req)) {
 731                                 result = PTR_ERR(req);
 732                                 break;
 733                         }
 734                         nfs_lock_request(req);
 735                         req->wb_index = pos >> PAGE_SHIFT;
 736                         req->wb_offset = pos & ~PAGE_MASK;
 737                         if (!nfs_pageio_add_request(desc, req)) {
 738                                 result = desc->pg_error;
 739                                 nfs_unlock_and_release_request(req);
 740                                 break;
 741                         }
 742                         pgbase = 0;
 743                         bytes -= req_len;
 744                         started += req_len;
 745                         user_addr += req_len;
 746                         pos += req_len;
 747                         count -= req_len;
 748                         dreq->bytes_left -= req_len;
 749                 }
 750                 /* The nfs_page now hold references to these pages */
 751                 nfs_direct_release_pages(pagevec, npages);
 752         } while (count != 0 && result >= 0);
 753
 754         kfree(pagevec);
 755
 756         if (started)
 757                 return started;
 758         return result < 0 ? (ssize_t) result : -EFAULT;
 759 }
 760
 761 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 762 {
 763         struct nfs_direct_req *dreq = hdr->dreq;
 764         struct nfs_commit_info cinfo;
 765         int bit = -1;
 766         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 767
 768         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 769                 goto out_put;
 770
 771         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 772
 773         spin_lock(&dreq->lock);
 774
 775         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 776                 dreq->flags = 0;
 777                 dreq->error = hdr->error;
 778         }
 779         if (dreq->error != 0)
 780                 bit = NFS_IOHDR_ERROR;
 781         else {
 782                 dreq->count += hdr->good_bytes;
 783                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 784                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 785                         bit = NFS_IOHDR_NEED_RESCHED;
 786                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 787                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 788                                 bit = NFS_IOHDR_NEED_RESCHED;
 789                         else if (dreq->flags == 0) {
 790                                 memcpy(&dreq->verf, hdr->verf,
 791                                        sizeof(dreq->verf));
 792                                 bit = NFS_IOHDR_NEED_COMMIT;
 793                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 794                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 795                                 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
 796                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 797                                         bit = NFS_IOHDR_NEED_RESCHED;
 798                                 } else
 799                                         bit = NFS_IOHDR_NEED_COMMIT;
 800                         }
 801                 }
 802         }
 803         spin_unlock(&dreq->lock);
 804
 805         while (!list_empty(&hdr->pages)) {
 806                 req = nfs_list_entry(hdr->pages.next);
 807                 nfs_list_remove_request(req);
 808                 switch (bit) {
 809                 case NFS_IOHDR_NEED_RESCHED:
 810                 case NFS_IOHDR_NEED_COMMIT:
 811                         kref_get(&req->wb_kref);
 812                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 813                 }
 814                 nfs_unlock_and_release_request(req);
 815         }
 816
 817 out_put:
 818         if (put_dreq(dreq))
 819                 nfs_direct_write_complete(dreq, hdr->inode);
 820         hdr->release(hdr);
 821 }
 822
 823 static void nfs_write_sync_pgio_error(struct list_head *head)
 824 {
 825         struct nfs_page *req;
 826
 827         while (!list_empty(head)) {
 828                 req = nfs_list_entry(head->next);
 829                 nfs_list_remove_request(req);
 830                 nfs_unlock_and_release_request(req);
 831         }
 832 }
 833
 834 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 835         .error_cleanup = nfs_write_sync_pgio_error,
 836         .init_hdr = nfs_direct_pgio_init,
 837         .completion = nfs_direct_write_completion,
 838 };
 839
 840 static ssize_t nfs_direct_do_schedule_write_iovec(
 841                 struct nfs_pageio_descriptor *desc, const struct iovec *iov,
 842                 unsigned long nr_segs, loff_t pos)
 843 {
 844         ssize_t result = -EINVAL;
 845         size_t requested_bytes = 0;
 846         unsigned long seg;
 847
 848         for (seg = 0; seg < nr_segs; seg++) {
 849                 const struct iovec *vec = &iov[seg];
 850                 result = nfs_direct_write_schedule_segment(desc, vec,
 851                                                            pos);
 852                 if (result < 0)
 853                         break;
 854                 requested_bytes += result;
 855                 if ((size_t)result < vec->iov_len)
 856                         break;
 857                 pos += vec->iov_len;
 858         }
 859
 860         if (requested_bytes)
 861                 return requested_bytes;
 862
 863         return result < 0 ? result : -EIO;
 864 }
 865
 866 #ifdef CONFIG_BLOCK
 867 static ssize_t nfs_direct_do_schedule_write_bvec(
 868                 struct nfs_pageio_descriptor *desc,
 869                 struct bio_vec *bvec, unsigned long nr_segs, loff_t pos)
 870 {
 871         struct nfs_direct_req *dreq = desc->pg_dreq;
 872         struct nfs_open_context *ctx = dreq->ctx;
 873         struct inode *inode = dreq->inode;
 874         ssize_t result = 0;
 875         size_t requested_bytes = 0;
 876         unsigned long seg;
 877         struct nfs_page *req;
 878         unsigned int req_len;
 879
 880         for (seg = 0; seg < nr_segs; seg++) {
 881                 req_len = bvec[seg].bv_len;
 882
 883                 req = nfs_create_request(ctx, inode, bvec[seg].bv_page,
 884                                          bvec[seg].bv_offset, req_len);
 885                 if (IS_ERR(req)) {
 886                         result = PTR_ERR(req);
 887                         break;
 888                 }
 889                 nfs_lock_request(req);
 890                 req->wb_index = pos >> PAGE_SHIFT;
 891                 req->wb_offset = pos & ~PAGE_MASK;
 892                 if (!nfs_pageio_add_request(desc, req)) {
 893                         result = desc->pg_error;
 894                         nfs_unlock_and_release_request(req);
 895                         break;
 896                 }
 897                 requested_bytes += req_len;
 898                 pos += req_len;
 899         }
 900
 901         if (requested_bytes)
 902                 return requested_bytes;
 903
 904         return result < 0 ? result : -EIO;
 905 }
 906 #endif /* CONFIG_BLOCK */
 907
 908 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq,
 909                                          struct iov_iter *iter, loff_t pos)
 910 {
 911         struct nfs_pageio_descriptor desc;
 912         struct inode *inode = dreq->inode;
 913         ssize_t result = 0;
 914
 915         NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
 916                               &nfs_direct_write_completion_ops);
 917         desc.pg_dreq = dreq;
 918         get_dreq(dreq);
 919         atomic_inc(&inode->i_dio_count);
 920
 921         NFS_I(dreq->inode)->write_io += iov_iter_count(iter);
 922
 923         if (iov_iter_has_iovec(iter)) {
 924                 result = nfs_direct_do_schedule_write_iovec(&desc,
 925                                 iov_iter_iovec(iter), iter->nr_segs, pos);
 926 #ifdef CONFIG_BLOCK
 927         } else if (iov_iter_has_bvec(iter)) {
 928                 result = nfs_direct_do_schedule_write_bvec(&desc,
 929                                 iov_iter_bvec(iter), iter->nr_segs, pos);
 930 #endif
 931         } else
 932                 BUG();
 933
 934         nfs_pageio_complete(&desc);
 935
 936         /*
 937          * If no bytes were started, return the error, and let the
 938          * generic layer handle the completion.
 939          */
 940         if (result < 0) {
 941                 inode_dio_done(inode);
 942                 nfs_direct_req_release(dreq);
 943                 return result;
 944         }
 945
 946         if (put_dreq(dreq))
 947                 nfs_direct_write_complete(dreq, dreq->inode);
 948         return 0;
 949 }
 950
 951 static ssize_t nfs_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 952                                 loff_t pos)
 953 {
 954         ssize_t result = -ENOMEM;
 955         struct inode *inode = iocb->ki_filp->f_mapping->host;
 956         struct nfs_direct_req *dreq;
 957         struct nfs_lock_context *l_ctx;
 958
 959         dreq = nfs_direct_req_alloc();
 960         if (!dreq)
 961                 goto out;
 962
 963         dreq->inode = inode;
 964         dreq->bytes_left = iov_iter_count(iter);
 965         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 966         l_ctx = nfs_get_lock_context(dreq->ctx);
 967         if (IS_ERR(l_ctx)) {
 968                 result = PTR_ERR(l_ctx);
 969                 goto out_release;
 970         }
 971         dreq->l_ctx = l_ctx;
 972         if (!is_sync_kiocb(iocb))
 973                 dreq->iocb = iocb;
 974
 975         result = nfs_direct_write_schedule(dreq, iter, pos);
 976         if (!result)
 977                 result = nfs_direct_wait(dreq);
 978 out_release:
 979         nfs_direct_req_release(dreq);
 980 out:
 981         return result;
 982 }
 983
 984 /**
 985  * nfs_file_direct_read - file direct read operation for NFS files
 986  * @iocb: target I/O control block
 987  * @iter: vector of buffers into which to read data
 988  * @pos: byte offset in file where reading starts
 989  *
 990  * We use this function for direct reads instead of calling
 991  * generic_file_read_iter() in order to avoid gfar's check to see if
 992  * the request starts before the end of the file.  For that check
 993  * to work, we must generate a GETATTR before each direct read, and
 994  * even then there is a window between the GETATTR and the subsequent
 995  * READ where the file size could change.  Our preference is simply
 996  * to do all reads the application wants, and the server will take
 997  * care of managing the end of file boundary.
 998  *
 999  * This function also eliminates unnecessarily updating the file's
1000  * atime locally, as the NFS server sets the file's atime, and this
1001  * client must read the updated atime from the server back into its
1002  * cache.
1003  */
1004 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
1005                              loff_t pos)
1006 {
1007         ssize_t retval = -EINVAL;
1008         struct file *file = iocb->ki_filp;
1009         struct address_space *mapping = file->f_mapping;
1010         size_t count;
1011
1012         count = iov_iter_count(iter);
1013         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
1014
1015         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
1016                 file, count, (long long) pos);
1017
1018         retval = 0;
1019         if (!count)
1020                 goto out;
1021
1022         retval = nfs_sync_mapping(mapping);
1023         if (retval)
1024                 goto out;
1025
1026         task_io_account_read(count);
1027
1028         retval = nfs_direct_read(iocb, iter, pos);
1029         if (retval > 0)
1030                 iocb->ki_pos = pos + retval;
1031
1032 out:
1033         return retval;
1034 }
1035
1036 /**
1037  * nfs_file_direct_write - file direct write operation for NFS files
1038  * @iocb: target I/O control block
1039  * @iter: vector of buffers from which to write data
1040  * @pos: byte offset in file where writing starts
1041  *
1042  * We use this function for direct writes instead of calling
1043  * generic_file_write_iter() in order to avoid taking the inode
1044  * semaphore and updating the i_size.  The NFS server will set
1045  * the new i_size and this client must read the updated size
1046  * back into its cache.  We let the server do generic write
1047  * parameter checking and report problems.
1048  *
1049  * We eliminate local atime updates, see direct read above.
1050  *
1051  * We avoid unnecessary page cache invalidations for normal cached
1052  * readers of this file.
1053  *
1054  * Note that O_APPEND is not supported for NFS direct writes, as there
1055  * is no atomic O_APPEND write facility in the NFS protocol.
1056  */
1057 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
1058                               loff_t pos)
1059 {
1060         ssize_t retval = -EINVAL;
1061         struct file *file = iocb->ki_filp;
1062         struct address_space *mapping = file->f_mapping;
1063         size_t count;
1064
1065         count = iov_iter_count(iter);
1066         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1067
1068         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1069                 file, count, (long long) pos);
1070
1071         retval = generic_write_checks(file, &pos, &count, 0);
1072         if (retval)
1073                 goto out;
1074
1075         retval = -EINVAL;
1076         if ((ssize_t) count < 0)
1077                 goto out;
1078         retval = 0;
1079         if (!count)
1080                 goto out;
1081
1082         retval = nfs_sync_mapping(mapping);
1083         if (retval)
1084                 goto out;
1085
1086         task_io_account_write(count);
1087
1088         retval = nfs_direct_write(iocb, iter, pos);
1089         if (retval > 0) {
1090                 struct inode *inode = mapping->host;
1091
1092                 iocb->ki_pos = pos + retval;
1093                 spin_lock(&inode->i_lock);
1094                 if (i_size_read(inode) < iocb->ki_pos)
1095                         i_size_write(inode, iocb->ki_pos);
1096                 spin_unlock(&inode->i_lock);
1097         }
1098 out:
1099         return retval;
1100 }
1101
1102 /**
1103  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1104  *
1105  */
1106 int __init nfs_init_directcache(void)
1107 {
1108         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1109                                                 sizeof(struct nfs_direct_req),
1110                                                 0, (SLAB_RECLAIM_ACCOUNT|
1111                                                         SLAB_MEM_SPREAD),
1112                                                 NULL);
1113         if (nfs_direct_cachep == NULL)
1114                 return -ENOMEM;
1115
1116         return 0;
1117 }
1118
1119 /**
1120  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1121  *
1122  */
1123 void nfs_destroy_directcache(void)
1124 {
1125         kmem_cache_destroy(nfs_direct_cachep);
1126 }