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