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